Fog-assisted Energy Efficient Cyber Physical System for Panic-based Evacuation during Disasters
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
© The British Computer Society 2021. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com doi: 10.1093/comjnl/bxaa201 Fog-assisted Energy Efficient Cyber Physical System for Panic-based Evacuation during Disasters Downloaded from https://academic.oup.com/comjnl/advance-article/doi/10.1093/comjnl/bxaa201/6249696 by guest on 17 August 2021 Sahil1,2, * and Sandeep Kumar Sood3 1 School of Computing, Indian Institute of Information Technology, Una, HP, India 2 Department of Computer Science and Engineering, Guru Nanak Dev University Regional Campus, Gurdaspur, PB, India 3 Department of Computer Applications, National Institute of Technology, Kurukshetra, HR, India ∗ Corresponding author: sahil.neelam@hotmail.com Disasters around the world have adversely affected every aspect of life and panic-health of stranded persons is one such category. An effective and on-time evacuation from disaster-affected areas can avoid any panic-related health problems of the stranded persons. Although the nature of disasters differ in terms of how they occur, the evacuation of stranded persons faces approximately same set of issues related to the communication, time-sensitive computation and energy efficiency of the devices operated in the disaster-affected areas. In this paper, a cyber physical system (CPS) is proposed that takes into account various challenges of the disaster evacuation, so an efficient on-time and orderly evacuation of stranded panicked persons could be realized. The system employs fog-assisted mobile and UAV devices for time-sensitive computation services, data relaying and energy-aware computation. The system uses a fog-assisted two-factor energy-aware computation approach using data reduction, which enables the energy-efficient data reception and transmission (DRecTrans) operations at the fog nodes and compensates to extend the period for other functionalities. The data reduction at fog devices employs Novel Events Identification (NEI) and Principal Component Anal- ysis (PCA) for detecting consecutive duplicate traffic and data summarization of high dimensional data, respectively. The proposed system operates in two spaces: physical and cyber. Physical space facilitates real-world data acquisition and information sharing with the concerned stakeholders (stranded persons, evacuation teams and medical professionals). The cyber space houses various data-analytics layers and comprises of two subspaces: fog and cloud. The fog space helps in providing real-time panic-health diagnostic and alert services and enables the optimized energy consumption of devices operate in disaster-affected areas, whereas the cloud space facilitates the monitoring and prediction of panic severity of the stranded persons, using a conditional probabilistic model and seasonal auto regression integrated moving average (SARIMA), respectively. Cloud space also facilitates the disaster mapping for converging the evacuation map to the actual situation of the disaster-affected area, and geographical population analysis (GPA) for the identification of the panic severity-based critical regions. The performance evaluation of the proposed CPS acknowledges its Logistic Regression-based panic-well being determination and real-time alert generation efficiency. The simulated implementation of NEI and PCA depicts the fog-assisted energy efficiency of the DRecTrans operations of the fog nodes. The performance evaluation of the proposed CPS also acknowledges the prediction efficiency of the SARIMA and disaster mapping accuracy through GPA. The proposed system also discusses a case study related to the pandemic disaster of coronavirus disease 2019 (COVID-19), where the system can help in panic-based selective testing of the persons, and preventing panic due to distressing period of COVID-19 outbreak. Keywords: Fog Computing; Panic Attacks; Unmanned Aerial Vehicle (UAV); Energy Efficiency; Logistic Regression; Principal Component Analysis (PCA); Seasonal Auto Regression Integrated Moving Average Section C: Computational Intelligence, Machine Learning and Data Analytics The Computer Journal, Vol. 00 No. 0, 2021
2 Sahil and S. Sood (SARIMA); Geographical Population Analysis (GPA); COVID-19; Internet of Things (IoT); Cloud Computing Received 3 April 2020; Revised 9 September 2020; Accepted 5 December 2020 Handling editor: Professor Gerard Parr 1. INTRODUCTION numbness or tingling, fear of dying and fear of losing Downloaded from https://academic.oup.com/comjnl/advance-article/doi/10.1093/comjnl/bxaa201/6249696 by guest on 17 August 2021 control or oncoming madness, whereas a clinical study [18] Disasters over the past several decades have led to considerable has found that the prevalence of panic attack is majorly destruction of physical infrastructure, a massive amount of manifested by the nine most commonly endorsed symptoms, human injuries, large-scale human causalities and substantial namely accelerated heart rate (98.3%), dizziness (96.0%), economic losses [1,2]. According to Swiss Re Institute [3], breathlessness (92.0%), sweating (88.0%), chest pain (85.0%), the natural and manmade disasters in 2019 have jointly caused chills (84.0%), trembling (84.0%), nausea (83.0%) and choking around 140 billion USD economic losses and claimed more (79.0%). than 11 000 lives. The Indian floods during the month of Panic-exposed situations can lead to the various medical con- August in 2019 have killed more than 420 persons and caused ditions [19] among persons viz. panic disorder, post-traumatic an economic loss of 7 billion USD [4]. Alone the disaster of stress disorder (PTSD), effect on cardiovascular system and wildfires in 2015 has victimized 494 000 persons and caused impairment of immune system. In panic disorder [20], the panic a 3.1 billion USD economic damage [5]. The hurricane Kat- attacks occur repeatedly and unexpectedly, and the person rina in 2005 has affected millions of people and killed more remains in a constant terror of having additional occurrences of than 1800, and caused 340 billion USD economical loss [6]. panic attacks. In PTSD [21], lasting consequences of traumatic Such substantial increase in number and impact of disasters ordeals viz. helplessness, intense fear etc. develop and remain in the recent past has affected every corner of the world, and for months. The long-term panic exposure can affect the human made disaster-associated risks the significant part of the threat- body, and increase the susceptibility of developing chronic space on this planet [7]. The disaster-associated risks can be medical conditions [22]. In a situation of panic, stress or mapped from the exacerbating effects of the climate change anxiety, the brain sends signals to fight or flee in response. The [8], and deteriorating situations like unplanned urbanization, body responds by releasing hormones like cortisol. However, demographic changes and unpreparedness to deal with disas- the long-term exposure of cortisol impairs the immune system, ter events [9]. Resultantly, the human population exposure to as cortisol prevents the release of chemicals which causes disasters has increased significantly [10] and posing various inflammation, which consequently affects the capability of the threats to the entire humankind. One such threat is panic-related immune system to protect the body against infections. That’s health problems. why persons with the chronic condition of panic disorder may The exposure to potentially traumatic events like disasters be likely more susceptible to catch flu, common cold or other has severe consequences in various health-related issues, and kinds of infections. In a situation of panic, breathing may this can be attributed from the extensive literature, which become shallow and rapid, a situation called hyperventilation, indicates that the high incidence of panic attacks (69–77%) where the body allows lungs to inhale more oxygen, so the occurs among the persons exposed to traumatic events and oxygen could transport in the body quickly, and prepare the causes various panic-related health problems [11] [12]. The body to respond to the situation of panic. The situation may impact of witnessing various traumatic events during disasters, become worse for a person suffering from Chronic Obstructive on stranded persons, can be characterized by the feeling of Pulmonary Disease (COPD). Panic situations can also accel- fear, which prompt stern physical reactions known as panic in erate the heart rate. A person suffering from vasoconstriction disasters [13]. These incidences of panic occurred in the form (a medical condition, where the blood vessels of the person of sudden periods of intense fear termed as panic attacks. The are narrow) may experience increased vulnerability of coronary sudden onset of a panic attack without any warning causes events, in a situation of panic. Henceforth, the consideration intense discomfort [14] [15], and typically spans over 10 to 20 of panic health of the stranded persons during the disaster is min, but may last for more than an hour in extreme cases [16]. critical for on-time and effective evacuation, and to avoid any The diagnostic criteria of DSM-V of American Psychiatric panic-related post-disaster health problems. Association [17] state that a panic attack can be characterized by the occurrence of at least four of the symptoms, namely 1.1. Challenges accelerated heart rate, breathlessness, chest pain, trembling, feeling of choking, abdominal discomfort or nausea, chills During disasters, the panic-health of the stranded persons is or feeling of warmth, insatiability or faintness, dizziness, impacted due to the witnessing of various dreadful situations. derealization or feeling unreal, sweating, hallucinations or The evacuation of such persons on a priority basis can prevent Section C: Computational Intelligence, Machine Learning and Data Analytics The Computer Journal, Vol. 00 No. 0, 2021
Fog-assisted Energy Efficient CPS for Evacuation 3 the ill-effects of panic on their health. However, the situations management [28], healthcare [29], transportation [30], indus- of critical infrastructure like roads, bridges, buildings etc., in trial manufacturing [31], and alike. The integration of IoT net- the disaster-affected areas, change over the time due to the work with cloud computing has provided these smart things and disruptions and dynamic phenomenon like the flooding of mobile devices higher storage and computation capabilities, roads, smoke, debris, etc. Hence, in such situations, the real- and made the system to analyze the physical world scenarios time monitoring and analysis of the stranded persons’ health for different domains remotely. The advent of fog computing and situational information of the disaster-affected areas can in Cloud-IoT scenario has further exploited the computation enable the evacuation process effective and timely [23]. This functionalities of the cyber space to provide shorter response, Downloaded from https://academic.oup.com/comjnl/advance-article/doi/10.1093/comjnl/bxaa201/6249696 by guest on 17 August 2021 can be only possible if the Information and Communication location-aware computation and less dependency on network Technologies (ICT) could complement the evacuation teams in bandwidth [32] [33]. The placement of fog-devices at the net- realizing the entire situation from remote locations, and enable work edge, near to data sources, and away from cloud servers, them to act accordingly. makes this paradigm suitable for time-sensitive applications. However, the various limitations restrict the realization of In situations like disasters, where entire communication the effective evacuation process: (1) communication, (2) time- infrastructure is highly affected and most of the time leads sensitive computation and (3) energy efficiency of the devices, to the complete breakdown, the aggregation of data from the in disaster-affected areas. Communication is one of the main sensors, and relay to the cloud servers, requires a kind of ad concerns in disaster-affected areas. The communication infras- hoc networking infrastructure. The UAVs facilitates various tructure is highly affected by the destructive nature of the such data-related activities in the disasters by acting as mobile disasters, and most of the time leads to the complete breakdown relay [34]. UAVs can reach in such hostile situations, and of the entire communication network [24]. In such situations, operate as relay nodes in facilitating the communication of the partial or complete failure of communication between the ground nodes with the remote data centers [35] [36] [37]. The stranded persons and response system can lead to the delays integration of these technological paradigms, and addressing and faults in response and subsequently to unavoidable loss of evacuation-related challenges can end up in an effective on- of life. In situations like disasters, where on-time decision- time and orderly evacuation system. making is critical for various time-sensitive applications, the incapability of evacuation teams and monitoring frameworks 1.3. Focus to monitor real-time dynamics viz. health parameters and envi- ronmental information can deteriorate the on-time and orderly In the times where the quality of life is quantified in terms evacuation of the stranded persons. The power grids or source of up to what extent a human can achieve the sustainability of power also get affected due to the destructive impact of the regarding health, living, handling of unexpected events and disasters and might result in the unavailability of power supply other matters, the technology plays an essential role in every in disaster-affected areas. Hence, the majority of devices in aspect of human life. The incorporation of ICT in the domain the post-disaster phase operates primarily on battery-sourced of disaster management is one such dimension, which focuses power [25]. These devices remain alive for a limited period of on the sustainability of human beings regarding the handling time, and need to be recharged. Hence, the operations of these of unexpected disaster events [38]. In this paper, a cyber phys- devices must be energy-aware to keep these devices alive for ical system (CPS) is proposed that takes into account various an elongated period. challenges of the disaster evacuation, so an effective on-time and orderly evacuation of stranded panicked persons can be realized. The CPS can be considered as the orchestration of 1.2. Motivations physical systems and distributed computing, where properties Evacuation is the key in the post-disaster management activities of the physical systems are acquired by the transducers and and attracting more and more attentions of the nations, industry analyzed by the computation resources of the cyber systems and academia around the world. The on-time and orderly evac- [39]. The proposed system operates in two spaces: physical uation of stranded persons from disaster-affected areas can save and cyber. Physical space facilitates the acquisition of various lives and reduce the destructive effects of disasters on human disaster-related attributes, using wireless body area network health effectively. However, for effective disaster evacuation, (WBAN)-assisted biosensors and behavioural sensors, and IoT- the real-time situation awareness viz. damages to the paths, assisted environmental sensors. It also provides alerts to the physical environment and monitoring of stranded persons is concerned stakeholders (stranded persons, evacuation teams crucial [26] and provides a holistic picture of the disaster- and medical professionals). The cyber space houses various affected areas to the evacuation teams. data-analytics layers and comprises of two subspaces: fog The advancements in sensor technology and wireless com- and cloud. The fog space employs local data analytics for munication, and their assimilation in the internet of things time-sensitive and energy-aware computation using fog com- (IoT) [27] has promoted the wide-scale deployment of smart puting and relays the acquired physical data to the remote things and mobile devices in various domains like disaster cloud servers in the cyber space. The fog space helps in Section C: Computational Intelligence, Machine Learning and Data Analytics The Computer Journal, Vol. 00 No. 0, 2021
4 Sahil and S. Sood providing real-time panic-health diagnostic and alert services, work lacks the corresponding feature. The feature EEA also and enables two-factor energy efficiency of fog devices oper- mentions the approaches which have been employed by the ated in disaster-affected areas using data reduction. It houses corresponding works for achieving energy efficiency. The three layers: synchronization (SYN), panic wellbeing determi- comparison depicts that none of the evacuation-based work nation and smart decision making (PWD) and energy conser- has considered the healthcare aspect during the process of vation layer (ECL). evacuation. The literature review has also not found any smart After processing at fog space, the data advance to the cloud healthcare framework regarding the panic attacks. Based on space for panic severity analysis and disaster mapping. The the analysis of the literature review and comparison of the Downloaded from https://academic.oup.com/comjnl/advance-article/doi/10.1093/comjnl/bxaa201/6249696 by guest on 17 August 2021 cloud space employs two layers, namely Panic Severity Analy- state-of-the-art works, the presented CPS has focused on sis (PSA) Layer and Disaster Mapping Layer (DML). The PSA the criticality of panic-based healthcare during disasters, and monitors and predicts the panic severity of the stranded per- proposed a panic-based evacuation system that enables on- sons using conditional probabilistic analysis and seasonal auto time and orderly evacuation of the stranded persons and also regression integrated moving average (SARIMA) respectively, considers the energy constraints of the devices operated in whereas the DML employs geographical population analysis disaster-affected areas. (GPA) to identify the critical regions and evacuation-priority This paper has been organized into five parts. The sec- within those regions based on the determined panic severity of ond part presents the proposed CPS and discusses its various the stranded persons. DML also converge the evacuation maps subsystems. The third part evaluates the performance of the to the actual situations, based on the monitored disaster status proposed system. The fourth part discusses a case study related of the routes. The knowledge of evacuation routes, identified to the coronavirus disease 2019 (COVID-19) based panic. The critical regions and panic severity-based evacuation priority of fifth part concludes the findings from the presented research. the stranded person in those regions provides the evacuation teams with the capability to plan their strategy for on-time and 2. PROPOSED SYSTEM orderly evacuation. The proposed CPS operates in two spaces: physical and cyber, 1.4. Contributions as shown in Fig.1. The physical space facilitates the process of data acquisition through various sensors and the cyber space The proposed CPS has a significant set of contributions in houses various data analytics layers. The cyber space comprises the domain of energy-aware disaster-oriented healthcare and of two subspaces, namely fog space and cloud space. Fog space evacuation as follows. employs fog computing-based local data analytics for time- sensitive and energy-aware computation at the fog nodes. It • This paper contributes a smart framework for panic- houses three layers: SYN, PWD and ECL. The acquired data based on-time and orderly evacuation of the stranded from physical space are processed at fog space, and finally persons. stored at the cloud storage in cloud space. The cloud space • This paper provides a fog-assisted two-factor energy employs two layers namely PSA Layer and DML, and facili- efficiency approach using data reduction. tates the analysis of panic severity of the stranded persons, and • This paper provides a data reduction approach names as disaster mapping by converging evacuation maps to the actual Novel Event Identification to identify two-dimensional situation, and identification of panic severity-based critical duplicate data. regions, and prioritizing of the evacuation of panicked persons • This paper provides an approach names as Geographi- in those regions. The proposed system facilitates on-time and cal Population Analysis to identify the disaster critical orderly evacuation of the stranded persons. Each space of the regions based on the panic severity of the stranded proposed CPS has been explained as follows. persons. 2.1. Physical Space 1.5. State-of-the-Art Literature and Paper Organization The responsibility of the physical space is to acquire the var- Table 1 depicts various significant state-of-the-art related ious disaster-related attributes and provide alerts to the con- works. The table represents the major contributions of these cerned stakeholders (stranded persons, evacuation teams and state-of-the-art works, and compare them based on nine feature medical professionals). This space collects the panic health viz. Fog/Edge Computing (FC/EC), Cloud Computing (CC), (PanHealth)-related attributes of the stranded persons, and the IoT, Real Time Monitoring (RTM), Prediction Modeling (PM), disaster environment (DisEnvi)-related attributes of the sur- Strategy Based Evacuation (SBE), Healthcare Aspect (HCA), roundings, as shown in Table 2. The proposed system employs Information Sharing √ (IS) and Energy Efficiency Approach biosensors, behavioral sensors and WBAN for acquiring Pan- (EEA). The tick ( ) indicates that the work comprises of Health attributes [18]. The PanHealth-related physiological the corresponding feature, and cross (×) indicates that the attributes like heart rate, breathlessness, chest pain, nausea and Section C: Computational Intelligence, Machine Learning and Data Analytics The Computer Journal, Vol. 00 No. 0, 2021
TABLE 1. State-of-the-art comparison. Authors Major contribution Year FC/EC CC IoT RTM PM SBE HCA IS EEA √ √ √ √ √ √ Xu et al. [6] Relationship-based 2018 × × × personalized evacuation scheme √ √ √ √ √ Bhattacharjee et al. [40] Crowdsensing-based evacuation 2019 × × × × maps building using smartphone-based delay tolerant networks √ √ √ √ Karthik and Suja [41] Wireless Sensor Network 2019 × × × × × (WSN)–based geographic map-oriented path discovery to multiple exits √ √ √ Fathy and Barnaghi [27] Energy-aware communication 2019 × × × × × Data Reduction through reduced data transmission √ √ √ Ejaz et al. [42] Coverage-area based UAV 2020 × × × × × Optimized data scheduling for energy efficient collecting route data collection √ √ √ √ √ Akmandor [43] Sensor-based machine learning 2018 × × × Data Reduction inference and compression √ √ √ √ √ √ Santamaria et al. [44] Cognitive Intelligence-based 2018 × × Data Reduction human activity using behavioral sensors √ √ √ √ √ √ Gia et al. [45] Physiological, behavioral and 2019 × × Adaptive Sampling environmental The Computer Journal, Vol. 00 No. 0, 2021 Fog-assisted Energy Efficient CPS for Evacuation attributes-based health monitoring √ √ √ √ √ √ √ Asghari et al. [46] Prediction of medical conditions 2019 × × for providing appropriate health services. √ √ √ √ √ √ √ √ Proposed CPS Fog-assisted energy efficient 2021 Data Reduction panic-based disaster evacuation Section C: Computational Intelligence, Machine Learning and Data Analytics 5 Downloaded from https://academic.oup.com/comjnl/advance-article/doi/10.1093/comjnl/bxaa201/6249696 by guest on 17 August 2021
6 Sahil and S. Sood Downloaded from https://academic.oup.com/comjnl/advance-article/doi/10.1093/comjnl/bxaa201/6249696 by guest on 17 August 2021 FIGURE 1. Architecture of panic-oriented evacuation-based cyber physical system. sweating are acquired by the wearable biosensors, whereas the PanHealth-related behavioral attributes like dizziness, chills, trembling and choking are acquired by the wearable behav- ioral sensors. The mobile devices of the stranded persons act as the sink nodes [47] in the WBAN, for aggregating Pan- Health attributes from the biosensors and behavioral sensors, in the physical space. The WBAN also acquires the location of the persons using GPS sensor of the sink nodes. The wifi- capability of the sink nodes provides the WBAN with an extension to integrate with the IoT network to transmit the acquired PanHealth attributes to the cyber space for advanced processing [48]. The IoT-assisted environmental sensors, present in ambient, acquire DisEnvi attributes, namely visibility range, temperature of the structures and environment, water level, smoke detection, FIGURE 2. Fog Network. tilt in structures and obstacle in the path. These environmen- tal sensors are implanted in the ambient viz. in-pavements, buildings, etc. These sensors transmit the location of their own placement along with acquired DisEnvi attributes in the IoT 2.2.1.Fog Space network. The UAVs act as the sink nodes in the IoT network Fog space employs fog computing-based local data analytics for aggregating DisEnvi attributes from environmental sensors, for time-sensitive and energy-aware computation at the fog in the physical space. The acquired data by mobile devices and nodes. It houses three layers: SYN, PWD and ECL. In this UAVs from Physical space are transmitted into cyber space for space, two types of fog nodes, User Fog Node (UFN) and various data analytical processing. Network Fog Node (NFN), operate at the different levels of the fog network, as shown in Fig. 2, and host three layers of fog space, as shown in Fig. 1. The acquired PanHealth data using WBAN sensors, from physical space, are transmitted to 2.2. Cyber Space the persons’ mobile devices. These devices act as low-level The acquired data from physical space arrive in the cyber fog nodes for PanHealth data and are called as UFNs in the space. The cyber space houses various data analytics layers fog network. These UFNs are present in the proximity of available at the different phases of the panic-based evacuation the persons and provide local data analytics for time-sensitive process. This space comprises of two subspaces, namely fog and energy-aware computation. The UFNs also act as smart space and cloud space. The explanation of each subspace is as gateways for WBAN, by housing three layers: SYN, PWD and follows. ECL, and extend the WBAN to the IoT network. Section C: Computational Intelligence, Machine Learning and Data Analytics The Computer Journal, Vol. 00 No. 0, 2021
Fog-assisted Energy Efficient CPS for Evacuation 7 TABLE 2. Disaster-related datasets. S. No. Dataset Description IoT technology Attributes 1. PanHealth Data about the health related Optical Heart Rate Sensors, Heart rate, Breathlessness, Chest pain, dataset physiological and behavioral ECG Sensors, Contraction Nausea, Sweating, Dizziness, Chills, attributes of the stranded Sensors, Capacitive Humidity Trembling, Choking, Location. persons. Sensors, Inertial Sensors, Accelerometer, Piezoelectric Downloaded from https://academic.oup.com/comjnl/advance-article/doi/10.1093/comjnl/bxaa201/6249696 by guest on 17 August 2021 Sensors, GPS Sensors. 2. DisEnvi dataset Environmental data regarding Infrared Sensors, Temperature Visibility, Temperature, Water level, the surroundings of the stranded Sensors, Ultrasonic Depth Smoke detection, Tilt in structures, persons. Sensors, Photoelectric Sensors, Obstacle range, Disaster Location. Electrolytic Sensors, GPS Sensors. The acquired DisEnvi data using IoT-assisted environmental 2.2.1.2.Panic Wellbeing Determination and Smart Decision sensors, from physical space, and the locally processed Pan- Making Layer Health data at UFNs in fog network, are transmitted to the The UFNs host this layer for local data analytics and real- UAVs. These UAVs act as sinks for DisEnvi data and as upper time panic wellbeing determination at the users’ premises. level fog nodes for UFNs processed PanHealth data in the This layer, based on the acquired PanHealth data, continuously fog network. These devices are called as NFNs in fog space. classifies the panic wellbeing (PW) of the stranded person in The NFNs fly in proximity to the environmental sensors and one of the two classes, namely Normal (NOR) or Abnormal mobile devices, for aggregating DisEnvi and PanHealth data, (ANOR). The class of NOR depicts that the PW of the persons and providing energy-aware computation to the aggregated is normal, and does not require any special consideration in data by housing two layers: SYN and ECL, as shown in Fig. 1. the process of evacuation for the current instance. The class The detailed explanation of each layer of the fog space is as of ANOR depicts that the PW of the person is abnormal, follows. means the person is panicked, and requires immediate medical guidance, and prioritized consideration further in the process of evacuation analytics. The PanHealth data of the stranded persons at a particular instance ti , in the proposed system, comprise of nine acquired attributes and forms a PanHealth 2.2.1.1.Synchronization Layer vector (PaHi ). The incidence of a panic attack is determined The synchronization layer acts as the entry point to the cyber in that PaHi , based on the identification of any four and more space, for the data acquired from physical space. This layer symptoms. Hence, the PWD layer employs logistic regression is present at both the UFNs and NFNs and performs the [50] for classifying the PW of the stranded persons. The logistic task of global synchronization of the acquired data. The regression fits for the categorical classifications, where deci- remotely deployed sensors have different internal clocks: non- sion boundaries are defined based on the threshold of various synchronized, synchronized and no clock [49], and the acquired scenarios, and the same is required in the determination of data from these sensors do not synchronize globally. Hence, PW of the stranded persons. The logistic regression defines the proposed CPS has programmed the gateway nodes, i.e. the decision boundaries as linear or nonlinear and classifies the UFNs in WBAN and NFNs in IoT network for synchronizing data into categorical classes. The logistic regression uses Eq. 1 the acquired PanHealth data and DisEnvi data, respectively. to classify the PanHealth records based on the threshold using These gateways globally synchronize the acquired data using a function, as follows. absolute global time stamps. These gateways tag the absolute global time instances on the values of the acquired attributes and synchronize the various events on the global timescale. eω0 +ω1 .PaHi The global synchronization of acquired data helps in depicting γ = (1) 1 + eω0 +ω1 .PaHi the holistic picture of the disaster-affected area. It facilitates various time-sensitive activities like real-time analysis of panic wellbeing, panic severity monitoring and prediction and where γ is the output the logistic regression, e is the base disaster mapping. The synchronized PanHealth data transmit to of natural logarithmic, ω0 is the intercept of bias, ω1 is the the UFN-hosted PWD layer, whereas the synchronized DisEnvi coefficient of the PanHealth record and PaHi is the PanHealth Data transmit to the NFN-hosted ECL, as shown in Fig. 1. record of the person at the time instance ti . Section C: Computational Intelligence, Machine Learning and Data Analytics The Computer Journal, Vol. 00 No. 0, 2021
8 Sahil and S. Sood The categorical classification requires the output of the in hostile situations like disasters. Since, the fog devices in response to be either 1 or 0, and the Bernoulli distribution the proposed system facilitate data collection, time-sensitive considers the probability of PW =1 (ANOR), if the output of local data analytics, data caching and transmission relay, their the function is γ , and correspondingly the probability of PW energy-consumption consideration is critical for the effective = 0 (NOR) if the output of the function is γ -1. But the linear operation of evacuation in disaster-hit areas, where the source relation of γ and PaHi violates the constraint of the probability of power regeneration is not available [27]. Hence, the energy- to range between 0 and 1. Hence, the logistic regression uses aware computation at the UFNs and NFNs can extend the sigmoid function, which provides s-shaped curve to classify lifetime of fog nodes and guarantee the quality of service. Downloaded from https://academic.oup.com/comjnl/advance-article/doi/10.1093/comjnl/bxaa201/6249696 by guest on 17 August 2021 the data into one of the category using the threshold of the In IoT networks, the data reception and transmission (DRec- output of the function, as shown in Eq. 2. Trans) poses as the dominant factor for the energy consumption of the devices, and consumes higher energy than the data 1 processing [50] [51]. However, by limiting the communica- PW = sig(γ ) = (2) 1 + e−γ tion between the devices, the DRecTrans energy consumption of the devices can be reduced [52], and compensated for where PW depicts the class of the determined panic wellbeing, extending the period for other functionalities. The ECL of and sig(γ ) is the categorized value of the logistic regression the proposed system considers these factors, responsible for output γ . energy consumption in fog nodes, and enables energy-efficient The fog-based local data analytics functionality at the UFN operations of the devices through energy-aware computation enables real-time PW determination and alert generation to using data reduction at the fog nodes. The ECL focuses on three concerned stakeholders. The alert generation provides on-time dimensions for energy-aware computation: data reduction, data diagnostic alerts and medical guidance in the event of a panic quality and energy conservation. The ECL addresses these attack, to the person and his/her relatives for immediate care, concerns by reducing data transmission from UFNs to NFN and as shown in Algorithm 1. The determined PW class, along with from NFN to cloud servers in such a manner that the quality acquired PanHealth record, is further transmitted to the ECL. of data can be retained by reconstructing the same data as if they were from the source, and the energy can be conserved Algorithm 1 Panic well-being determination and alert through reduced DRecTrans operations. The ECL employs the generation following energy model to determine the energy consumption of UFNs and NFNs. Input: PanHealth Record PaHi , time instance ti 1: Until the person get evacuated Energy Model 2: Determine the current time stamp ti The energy model considers the energy consumption of fog 3: Map the PaHi to the feature space nodes during the operations of data collection, local data ana- 4: Determine the PW of the mapped sample using, lytics, data caching and data transmission. The model denotes eω0 +ω1 .PaHi 5: γ = 1+e ω0 +ω1 .PaHi the energy consumption for data collection as ER , local data 6: PW = sig(γ ) = 1+e1−γ analytics and data caching as EP and data transmission as ET . 7: Send Diagnostic alert (PW) to the person The proposed energy model refers to the energy consump- 8: If PW == ANOR, then tion characteristics of the energy model [53] [54]. The model 9: Send Guidance alert to person & relatives considers the energy consumption of a UFN for collecting a 10: Transfer (PaHi ∪ PW) to the ECL single data value as α Joules. Suppose, a UFN receives m 11: Exit data values in an instance, then the total energy consumption (ER _UFN ) for collecting m data values is shown Eq. 3. The Output: PW of the stranded person and generated alerts model considers the energy consumption of a UFN for locally analyzing and caching a single data value as β Joules. Suppose, 2.2.1.3.Energy Conservation Layer a UFN processes and caches m data values, then the total During disasters, the infrastructure and essential services are energy consumption (EP _UFN ) for analyzing and caching m highly affected by the destructive nature of the disasters, and data values is shown in Eq. 4. The model considers the energy most of the time leads to the complete breakdown. In such consumption of a UFN for transmitting a single data value hostile situations, the majority of disaster management oper- as η Joules. Suppose, the UFN transmits m’ data values after ations primarily operate on battery-powered devices. These processing, then the total energy consumption (ET _UFN ) for devices remain alive for a limited period of time, and need to transmitting m’ data values is shown in Eq. 5. be recharged. In the proposed system, the fog devices UFNs and NFNs operate in the disaster-affected areas. Therefore, the inefficient energy consumption in these devices could have a negative impact on the power-constrained fog operations ER _UFN = m ∗ α Joules (3) Section C: Computational Intelligence, Machine Learning and Data Analytics The Computer Journal, Vol. 00 No. 0, 2021
Fog-assisted Energy Efficient CPS for Evacuation 9 EP _UFN = m ∗ β Joules (4) Then, the total energy consumption (ET _NFN ) of an NFN for transmitting n’ data packets is shown in Eq. 13. −→ h (11) ET _UFN = m ∗ η Joules (5) = ρ ∗ h Joules (12) n Based on Equations 3–5, the total energy consumption of a ET _NFN = i Joules (13) Downloaded from https://academic.oup.com/comjnl/advance-article/doi/10.1093/comjnl/bxaa201/6249696 by guest on 17 August 2021 UFN (EUFN ) for relaying acquired data from sensors to NFN, i=1 for a particular time instance is shown in Eq. 6. The ECL at the UFN works on optimizing the EUFN by employing a mod- Based on Equations 9–13, the total energy consumption ule named Novel Event Identification (NEI), which reduces (ENFN ) of an NFN for relaying aggregated data from UFNs and the acquired data from sensors, so the transmission energy environmental sensors, to the cloud server for a particular time consumption ET _UFN can be minimized and resultantly could instance is shown in Eq. 14. optimize the overall EUFN . ENFN = ER _NFN + EP _NFN + ET _NFN Joules (14) EUFN = (m ∗ α) + (m ∗ β) + (m ∗ η) Joules (6) The ECL at the NFN works on optimizing the ENFN by The energy model further considers the data received from employing a module named Dimensionality Reduction, which UFNs and environmental sensors, to NFN, in the form of reduces the features of aggregated data, so the transmission packets, which may have variable length. The proposed energy energy consumption (ET _NFN ) can be minimized. Even the model only considers the number of data values, from a source energy consumption during the data collection (ER _NFN ) at (UFN or environmental sensor) as the length of the packet. NFN is also optimized due to the energy-aware computation During the relaying of data to the NFN, the model considers of data by NEI. Hence, the entire resultant of NEI and Dimen- n number of UFNs and environmental sensors, which transmit sionality reduction helps in minimizing the ENFN . Each module the data in an instance to that NFN. For collecting each packet, of this layer on UFNs and NFN is explained as follows. the energy consumption at the NFN depends upon the size of the packet. Let say a packet is of variable length h. Then, the 2.2.1.3.1.Novel Event Identification energy consumption () for collecting each packet is directly The ECL employs a module named Novel Event Identifica- proportion to the length of packet as shown in Equations 7 and tion (NEI) at UFNs to avoid consecutive duplicate PanHealth 8, and the total energy consumption (ER _NFN ) for collecting n data transmission to the NFN. The WBAN in physical space data packets at NFN is shown in Eq. 9. continuously acquires the PanHealth events, and UFNs in fog space continually monitor the panic wellbeing. However, the −→ h (7) PanHealth attributes may not change or remain the same, depend upon the body and surrounding events of the stranded = ρ ∗ h Joules (8) person. In such a scenario, the transmission of data to the NFN may involve the consecutive duplicate events, and even where ρ is the constant, which accounts for the increasing consecutive duplicate attribute values too. This can result in energy consumption with increase in packet length and vice- energy consumption for transmitting the duplicate data. The versa. energy of the UFNs can be conserved by avoiding such dupli- n cate data transmissions to the NFN. Hence, the NEI focuses on the identification of duplicate data, i.e. consecutive duplicate ER _NFN = i Joules (9) i=1 events and consecutive duplicate attribute values, and allows only novel events and novel attribute values to relay further in The model considers the energy consumption of an NFN for the fog space. locally caching, and processing a single packet as τ Joules. The sink node of the WBAN arranges the PanHealth The total energy consumption (EP _NFN ) for locally caching attributes acquired at the time instance ti , in a well-defined and processing the received and filled n’ data packets at NFN sequence to form a PanHealth vector (PaHi ), and follows is shown in Eq. 10. the same sequence throughout the process. The PWD layer classifies the PanHealth event PaHi at ti into one of the PW EP _NFN = n ∗ τ Joules (10) class and appends the classified PW value to the PaHi . This PaHi is analyzed by the NEI to identify the novel PanHealth The model considers the energy consumption of an NFN for events, and attribute values. The NEI creates a checkpoint in the transmitting a single packet as ’ Joules, which depends upon time-space, when a novel event is encountered, stores it in the the length of the packet h’, is shown in Equations 11 and12. log memory of the UFN, and transmits the event to the NFN. Section C: Computational Intelligence, Machine Learning and Data Analytics The Computer Journal, Vol. 00 No. 0, 2021
10 Sahil and S. Sood The NEI determines the duplication of the subsequent events 4: Create a checkpoint at ti one by one, by determining the fully matched event, fully 5: Flush log memory of UFN unique event and partially matched event, using a piecewise 6: uRec = PaHi match function as shown in Eq. 15. 7: Save uRec in log If the match function determines that the subsequent event, 8: If(isempty(binSeq[ti ])==0), then i.e. PaHi+1 is fully matched, the NEI does not transmit the 9: PaHCompi =omitFeatures(PaHi , binSeq[ti ]) subsequent event and continues to determine the matching 10: PaHCompi = PaHCompi ∪ binSeq[ti ] between the unique event and next subsequent events. The 11: Send PaHCompi to NFN Downloaded from https://academic.oup.com/comjnl/advance-article/doi/10.1093/comjnl/bxaa201/6249696 by guest on 17 August 2021 UFN uses cache memory to store the most recent event. If 12: Else (Send PaHi to NFN) NEI determines any subsequent event fully unique, it creates 13: mRec = PaHi+1 a new checkpoint to the time instances of the newly iden- 14: Save mRec in log tified novel event, flushes the content of the log memory, 15: If(match(uRec,mRec)==uRec==mRec), then saves the newly identified novel event in the log memory and 16: i++ transmits that event to the NFN. If the NEI determines that 17: Goto Step 13 the subsequent event partially matches with the unique event, 18: ElseIf(match(uRec,mRec)==∅), then NEI constructs a binary sequence of the partially matched 19: i++ subsequent event using function constructBinSeq that employs AND operation to determine the attribute values of the partially 20: Goto Step 4 matched event, which are different from the unique event, and 21: Else saves the binary sequence in binSeq. Based on the binSeq value, 22: i++ the function omitFeatures omits the matching values of the 23: binSeq[ti ]=constructBinSeq(uRec,mRec) partially matched event and makes the compressed form of the 24: Goto Step 4 partially matched event, i.e. PaHCompi . The NEI appends the 25: Exit corresponding binSeq with the compressed record (PaHCompi ) Output: Novel Data Transmission and sends it to the NFN. In this manner, the NEI also avoids the subsequent duplicate attribute values of the PanHealth data. The entire working of NEI is shown in Algorithm 2. Here, in 2.2.1.3.2. Dimensionality Reduction Algorithm 2, uReC and mRec signify the unique PanHealth The NFN acts as the ultimate sink for data from various record and subsequent matching PanHealth record (which is sources viz. UFNs and environmental sensors. The data arrived matching with the unique record), respectively. at the NFN includes the temporal PanHealth and DisEnvi match(PaH i , PaH i+1 ) events of the various stranded persons and various locations, ⎧ ⎫ respectively, which make the acquired data at the NFN a ⎪ ⎪ ∅ (PaH i ∩ PaH i+1 ) = PaH i = PaH i+1⎪ ⎪ high dimensional data. The high dimensional data present ⎪ ⎨PaH ⎪ ⎬ i+1 (PaH i ∩ PaH i+1 ) = ∅ various challenges like intense computation requirements, and = ⎪ ⎪ PaH − (PaH ∩ PaH ) ⎪ ⎪ increased error rate during analysis [55]. The transmission ⎪ ⎩ i i i+1 ⎪ ⎭ +binSeq (PaH i ∩ PaH i+1 ) ⊂ PaH i of such high-dimensional big data requires significant energy (15) consumption at the NFN. The ECL employs Dimensionality Reduction (DR) module at the NFN, which addresses the issue The function in Eq. 15 matches the subsequent PanHealth of energy conservation at the NFN, by employing energy- event PaHi+1 with the saved unique event PaHi , and only trans- aware computation using data summarization. However, the mits the outcome of the function based on three different crite- NEI-processed data from the UFN have omitted consecutive ria for fully matched, fully unique and partially matched event, duplicate events and values, and the DR module requires the respectively. The outcome of the match function determines the entire data for identifying the data patterns, and trends for data value of m’ (refer Eq. 5), such that m’ ≤ m. In this manner, summarization. That is why the DR has a filler component the transmission energy consumption of the UFN (ET _UFN ) (FL), which analyzes the collected data against the time-series minimizes and resultantly, the overall energy consumption of to fill the missing values in the received data. It replicates the the UFN (EUFN ) optimizes, by reducing the data transmission preceding value of the attributes in the missing value position, from UFN to NFN. in the time series. The DR module uses Principal Component Analysis (PCA) Algorithm 2 Novel Event Identification [56], which transforms the high-dimensional data, say Qn×d having n records and d dimensions into low dimensional sub- Input: Temporal PanHealth events space, say Q’n×d having n records and d’ dimensions such 1: Set Time counter i as 0 that d’ < d. PCA transforms the data in such a way that 2: Set binSeq[ ] as ∅ the data can be represented maximally using few dimensions. 3: For every temporal PanHealth event PaHi , The PCA identifies those dimensions or principal components Section C: Computational Intelligence, Machine Learning and Data Analytics The Computer Journal, Vol. 00 No. 0, 2021
Fog-assisted Energy Efficient CPS for Evacuation 11 (PC), which can retain maximum information of the common structure that exists in a dataset by using the concept of co- Algorithm 3 Dimensionality Reduction variance and Eigen’s (values and corresponding vectors), as Input: High dimensional dataset Qn×d shown in Equations 16 and 17. The transformation of data using Eigen vectors identify the PCs. The identified PCs or Eigen 1: Column standardization of dataset Q vectors depict the direction in which maximum variance of the 2: Determine the co-variance matrix S of column standard- data is retained, using Eigen values, as shown in Eq. 18. The ized dataset Q identified PCs provide the loading factors of each dimension to 3: Determine the Eigen values and corresponding Eigen Downloaded from https://academic.oup.com/comjnl/advance-article/doi/10.1093/comjnl/bxaa201/6249696 by guest on 17 August 2021 project the data values on those PCs, as shown in Eq. 19. The vectors of S using, PCs depict the maximum variance or spread of the data and are 4: λi .vi = S.vi , where where, i i:1−→d mutually uncorrelated. 5: Choose PCs account for maximum variance using cumu- lative variance S = QT Q (16) 6: Compute xij ’ of the PCj using loading factors as, 7: xij = xTi ∗ PCj where, S is the co-variance matrix, and QT is the transpose of 8: Exit the column standardized dataset Qn×d . Output: Reduced Dimensional dataset Q’n×d The column standardization of the dataset Q helps in moving the data points to the origin along with retaining the original The employed NEI reduces the packet length (m’), such spread of data, and the co-variance matrix analyzes the co- that m’ ≤ m, of the transmitting packet from the UFN; as dependence between the dimensions. a result, the energy consumption for collecting packets from UFNs reduces at NFN and subsequently optimizes ER _NFN . On λi .vi = S.vi (17) the other side, employed dimensionality reduction at the NFN reduces dimensions of data from d to d’, and results in reducing where, λi i:1−→d are Eigen values for of the d-dimensions of the energy consumption for transmitting a single packet (’), the co-variance matrix, and vi i:1−→d are the corresponding as the length of packet reduces, and subsequently optimizes Eigen vectors, such that λ1 ≥ λ2 ≥ λ3 ≥ . . . λd . The values ET _NFN . In this manner, the overall energy consumption of of λi and vi are determined by satisfying Eq. 17. λi s show the NFN reduces with the deployment of NEI at UFNs and the variance of the data retained by the corresponding Eigen dimensionality reduction at the NFN. vectors vi s. The PCA analyzes the variance or spread of data on a direction vi using corresponding Eigen values as show in 2.2.2.Cloud Space Eq. 18. The acquired data from physical space are processed at fog space, and finally stored at the cloud storage in the form of λi time-series data. The time series data consist of successive variance(λi ) = d (18) i=1 (λi ) observations made over the time interval [57] [58]. The cloud space employs two layers, namely PSA Layer and DML. Each PCA chooses the PCs which cover maximal variation using layer has been explained as follows. cumulative variance. Based on the identified PCs, the reduced dataset having chosen PCs as features of the new transformed 2.2.2.1.Panic Severity Analysis Layer data Q’n×d is transmitted to the remote servers in the cloud Panic Severity Analysis (PSA) layer analyzes the time-series space. The PCA computes PCi ∈ IRd , which represents the PanHealth-DisEnvi data of the panicked persons, for monitor- matrix of loading values corresponding to d-dimensions of Q. ing and predicting their panic severity in the form of Panic Using the loading values, and actual values of the attributes Severity Index (PSI). The PSI provides a probabilistic measure of the record (xi ), the transformed value (xi ’) is determined for analyzing the effects of the occurrence of the PanHealth- as shown in Eq. 19. Eq. 19 can be expanded, as shown and DisEnvi-related adverse events on the panic health of in Eq. 20. Algorithm 3 depicts the entire working of PCA the stranded persons and helps in facilitating the on-time and at NFN. orderly evacuation of stranded persons. A higher value of PSI indicates the possibility of severe panic attacks. Hence, the PSI xij = xTi ∗ PCj (19) is monitored and predicted for identifying the critical regions, and evacuation priority of people in those regions. The PSI is monitored in the form of conditional probability, as shown in Eq. 21. xij = xid1 ∗ PCjd1 + xid2 ∗ PCjd2 + xid3 ∗ PCjd3 + . . . PW PSI = P (21) + xidd ∗ PCjdd (20) e1 ∪ e2 ∪ e3 ∪ . . . ei Section C: Computational Intelligence, Machine Learning and Data Analytics The Computer Journal, Vol. 00 No. 0, 2021
12 Sahil and S. Sood Here, PW denotes the panic health class of the stranded MA terms, P is the order of seasonal AR (SAR) terms, D is person, and ei denotes the occurrence of an adverse event in the order of differencing or power of (1-LV ), A is the order a particular time instance. The PSI helps the cloud servers in of seasonal MA (SMA) terms and V is the seasonality period. monitoring the panic health severity of the stranded panicked The terms P (LV ) and P (LV ) converge the entire seasonal persons. Based on the current and past monitored PSI, the prediction function as shown in Equations 25 and 26. PSA layer further employs SARIMA prediction model for predicting the PSI. SARIMA is the extension to the prediction V )=1− V − 2V − 3V − ··· − PV P (L 1L 2L 3L PL model ARIMA, and considers the seasonal characteristics of Downloaded from https://academic.oup.com/comjnl/advance-article/doi/10.1093/comjnl/bxaa201/6249696 by guest on 17 August 2021 (25) the data using hyper-parameters for accounting the seasonality of the data series. The hyper-parameters in the SARIMA are similar to the parameters of the ARIMA; however, the hyper- parameters involve seasonal lag (i.e. minutes, hourly, daily, A (LV ) = 1 − 1 LV − 2 L2V − 3 L3V − · · · − A LAV weekly, monthly, yearly), which is specified by the seasonality (26) period variable. ARIMA employs auto regression (AR) and moving average (MA) to predict the value in a time-series based on the linear combination of present and past values, The PSI prediction using SARIMA involves four phases and prediction errors, respectively. ARIMA(p, d, a) is used to [59]: (I) Model Identification, (II) Parameter Identification, predict a value in time-series, as shown in Eq. 22. (III) Diagnostic Checking and (IV) Prediction. In model identi- fication phase, the time plot of the data is inspected for station- ary data and determines the value of d, so the variance of data Y t = φ 1 Y t−1 + φ 2 Y t−2 + · · · + φ p Y t−p + Z t + θ 1 Z t−1 + can be stabilized. After the identification of d, the preliminary θ 2 Z t−2 + · · · + θ a Z t−a (22) values, i.e. p, a, P, D and A are identified in this phase using Partial Autocorrelation Function (PACF) and Autocorrelation where Yt denotes the stationary data value at tth instance, for Function (ACF). the non-stationary data value St , using the differencing process The PACF determines the required order of AR terms i.e. of order d, such that Yt = St - St−d . St denotes the predicted PSI p, whereas the ACF depicts the amount of linear dependence of a person. p and a denote the number of AR terms and MA between data values of time series, which are separated by a terms, respectively. φ and θ are the AR and MA coefficients, lag of a. In parameter identification phase, the parameter and respectively. Yt denotes the predicted value, and Y t−1 . . . Y t−p corresponding standard errors are estimated using statistical denotes the previous p predicted data values. Zt denotes random measures: lease square estimation (LSE), maximum likeli- error for predicted data value and Z t−1 . . . Z t−a denote the hood (ML) and Yule–Walker. In diagnostic checking, different previous a prediction errors. The Eq. 22 can be represented models are evaluated, and their residuals are analyzed. The using lag operator (L) as shown in Eq. 23. model, which has the least residual or values of Mean absolute Error (MAE), Mean square error (MSE) and Root mean square φ p (L)Y t = θ a (L)Z t (23) error (RMSE), fits well and is selected. In prediction phase, the model predicts the PSI based on the fitted model. The where PSI prediction using SARIMA is illustrated in Algorithm 4. φ p (L)Y t = (1 − φ 1 L − φ 2 L2 − φ 3 L3 . . . φ p Lp )Y t The monitored and predicted PSI of the stranded persons are Lp .Y t = Y t−p transmitted to the DML, which analyzes the PSI time-series θ a (L)Z t = (1 − θ 1 L − θ 2 L2 − θ 3 L3 . . . θ a La )Z t to ascertain the highest panic severity of the person for a La .Z t = Z t−p particular future time-frame (ranges from current instance to The SARIMA predicts the PSI based on the past values of a particular future instance), so panicked critical regions and the PSI and past predicted errors. Since the PSI is based on evacuation priority of the stranded persons in those regions the number of symptoms appeared, and the duration of the could be identified. panic attacks, it is significant to consider the periodicity or seasonality of the time series in ARIMA prediction. Hence, the Algorithm 4 Panic Severity Index Prediction PSA layer employs SARIMA(p,d,a)(P,D,A)[V] to predict the Input: Monitored PSI Time-series data PSI of a stranded panicked person, as shown in Eq. 24 1: Determine the differencing values for stabilizing the φ p (L) P (L V )(1 − L)d (1 − LV )D St = θ a (L)A (LV )Z t (24) data. 2: Examine PACF, and ACF Plots to decide the structure of where and are seasonal AR and seasonal MA coefficients, SARIMA respectively. p is the order of nonseasonal AR terms, d is the 3: Determine the AR and MA coefficient for both non- order of nonseasonal differencing, a is the order of nonseasonal seasonal and seasonal Section C: Computational Intelligence, Machine Learning and Data Analytics The Computer Journal, Vol. 00 No. 0, 2021
You can also read