4.1.1 Fundamental Terms in IIoT

The different domains interconnected in IIoT are explained in the following.

4.1.1.4 Internet of Things

Internet of Things (IoT) is an organization of interconnected physical/virtual devices/objects which are given Unique IDentifiers (UIDs) and all are able to communicate over a network without the need of either Man-to-Man or Man-to-computer interaction. Smart vehicles, smart watches, home automation, and remote healthcare monitoring are some of the real-time examples of IoT applications. The working of IoT devices along with analysis is represented in Figure 4.2. All the data produced by IoT devices are not having the analytical power. The data may be noisy and heterogeneous. To solve this issue, preprocessing is required. After pre-processing, big data analytics can be combined with IoT devices to identify the hidden or unobserved patterns and innovate into useful information [7]. This will be helpful for the companies, organizations, or individuals to manage the large amounts of data and extract useful information by big data-mining to make important decisions, and to identify the trends and make predictions.

4.1.1.7 Machine Learning

ML is one of the sub domains of AI; it focuses on developing a set of algorithms to make the components or computing devices to learn automatically and improve their performance based on the experience (given as data samples). Algorithms used in ML are iterative to discover the hidden or natural patterns available in the data samples and use it for reliable and better predictions or decisions. After training, when these algorithms are exposed to unseen data, they are able to adapt independently.

ML is one of the most trending method and used almost in all the domains with the intersection of more technologies such as AI and data science. There are many emerging algorithms are being developed to process the data faster and make accurate predictions [12]. The applications of ML are tremendous in speech recognition, Natural Language Programming (NLP), healthcare applications, financial modeling, recommendation systems, face detection, and object recognition (Computer Vision) etc.

Learning phase of ML contains the following steps:

• (i) Collection of training data,
• (ii) Pre-processing,
• (iii) Feature selection or extraction,
• (iv) Model training with selected features.

There are three major kinds of ML algorithms and given in Figure 4.3.

• (i) Supervised learning,
• (ii) Unsupervised learning, and
• (iii) Reinforcement learning.

(i) Supervised Learning

Supervised learning has n-number of data samples as input-output pair (labeled data). The learning task is to find the function of mapping from input to output. Input is a vector and output is a single value/categorial label. If the prediction is categorial label, then it is called as classification, and in case of continuous value, it is called as regression. The mapping function may be linear or non-linear. Some of the supervised learning algorithms are Decision Trees (DT), Support Vector Machines (SVM), linear regression, Naive Bayes, logistic regression, K-Nearest Neighbor algorithm (k-NN for classification) and Neural Networks (Multilayer perceptron). Based on the algorithm selected, the parameters must be learned to improve the prediction accuracy. Parameters are significant in ML algorithms. In semi-supervised learning, most of the data samples are labeled and few are unlabeled.

Nowadays, Deep Learning (DL) algorithms are more popular, because it solves complex problems even though the data samples are varying, unstructured, and also interconnected. DL has multiple layers of neuron (similar to Artificial Neural Networks), along with one input layer, one output layer and multiple number of hidden layers are available. Algorithms like Back Propagation (BP) are used to learn the weights of interconnected neurons automatically by using the inputs such as images, text, or sound. There is no need of feature extraction because of convolution operation and pooling. Convolution does repeated filtering on input to get the feature map. Pooling operation makes down sampling and also dimensionality reduction.

(ii) Unsupervised Learning

In unsupervised learning, the data set is not labeled. The hidden pattern among the given data is identified and based on the similarities of features, and the given datasets are grouped into multiple numbers of clusters. Then, it may be labeled as Group 1, 2, …, n. There are different kinds of distance measures such as Manhattan distance, city block distance, and Euclidean distance, which are used to find the similarities between data samples. The usage of unsupervised learning is exploratory analysis and dimensionality reduction.

Figure 4.4 illustrates the difference between supervised and unsupervised learning using the same kind of data samples.

Fuzzy and K-means clustering are most commonly used for clustering. The objective of clustering is to reduce the dissimilarities within a cluster and increase the dissimilarities in inter-clusters. Dimensionality reduction reduces the feature set in the ML algorithms to take away only the useful and relevant features either for classification or clustering.

(iii) Reinforcement Learning

Reinforcement Learning (RL) is a goal-oriented approach and has multiple agents who interact with the real-world environment to make decisions. It may use either supervised or unsupervised learning to build the model. The agent reacts to the environment. It may be a hardware, software, or combination of them. When correct decision is made, the reward (positive score) will be given, in case of incorrect decision, penalty (negative score) will be given. The ultimate aim of RL is to get the highest cumulative score. In the nonappearance of a training sample-dataset, it is destined to learn from its experience. In robotics, RL behaves novel by using “Trial-and-Error” concept.

The elements of RL are given in the following:

• (i) Policy: It is a way of agent’s behavior based on the current situation. It is like stimulus-response.
• (ii) Reward function: Increasing reward is the goal of RL. If the action taken in the current state is toward the goal, then a positive value is awarded and called as award.
• (iii) Value function: It defines how much award the particular agent can get from the current state to the goal state. It is the impression of the behavior of the environment.
• (iv) Model of the environment: Model predicts the next state and corresponding reward. The working of a simple agent is represented in Figure 4.5.

4.1.2 Intelligent Analytics

As increasing numbers of connected sensors are used, they create voluminous data periodically. They should be aggregated before sending to remote server. To avoid the redundant and non-useful data, big data mining can be used. This helps the headache of data transmission.

Intelligent analytics helps to understand, analyze the data, and identify useful information which will be helpful to the organizations for making well-formed decisions and inform to the respective people in the organization at right time. It uses the concepts of big data, AI, and data science. Data visualization from different view point is to be taken care based on the needs of an application. By using appropriate tools, collection of current data, and history of information, analytics should result in accurate results. The mature predictions lead to optimization.

The intelligent analytics are categorized into three types. They are descriptive, predictive, and perspective using the mining, AI, and statistics [15].

• (i) Descriptive analytics: used in streaming the context. It will answer for the following queries: “What and why a particular event has happened?” and “what is happening now?”
• (ii) Predictive analytics: used to replying for the following queries: “What and why some event will happen in the future?”
• (iii) Prescriptive analytics: answering for the following queries “What and why should be done now”. It is attracting to the researchers, because it is a matured data analytics and lead to reach the optimal solutions.

4.1.3 Predictive Maintenance

The integration of sensed data with predictions is used in IIoT applications to find the significant relationships with predictions and the procedures of maintenance. In manufacturing industries, the objective is to assure the maximum throughput in the manufacturing line and to improve the production with reduction of cost and throughout the life cycle of an equipment. The data driven approaches are used to make the predictive maintenance. In any organization, each equipment is an important asset. To avoid the complete shut-down, prediction of faults is vital. Complete monitoring of each and every device may be useful to predict about when the servicing is required to avoid the fault and decide on when to back up the data. There are three kinds of maintenance procedures used in industries [16].

• (i) Predictive maintenance: where the component is restored or substituted after wear, fault, or break down.
• (ii) Preventive maintenance: customs sensor data for continuous monitoring of a system and evaluates it in contradiction with historical events to predict the failure before it happens.
• (iii) Corrective maintenance: recognizes the faults by a result of walkthroughs or inspections.

Based on the specific needs of applications, the guidance can be designed about how to utilize the streaming technologies and big data analytics tools for the use cases of predictive maintenance [17]. Each domain and application has different requirements, based on the requirements, the suitable analytical tool is chosen. While automating the complex operations, one of the challenges is reliability and safety of devices and data [18] which may have great impact on decision-making.

In this chapter, predictive maintenance is taken for assets safety during the hazardous events.

4.1.4 Disaster Predication and Safety Management

Natural disaster is beyond the human being’s control and unavoidable. Many countries faced different kinds of disasters and some are very deadly.

World Health Organization (WHO) stated that during the year 1900 to 2018, nearly 14 million disasters happened [19] in world wide.

4.1.5 Optimization

For any problem solving, multiple feasible solutions may be available. In each solution, it has different impact on the real-time scenario. Some of the nature-inspired optimization algorithms are Genetic algorithms (GA), Particle Swarm Optimization (PSO), Ant Colony Optimization (ACO), and Memetic Algorithms (MA). Selecting proper method improves the effectiveness. Efficient forecasting and scheduling operations play a vital role in the relief of saving people and lowering the damages in the event of disasters. These operations come under evolutionary and optimization problems. Response phase needs the concept of optimization to make selection of the location, allotment of medical tents, and evacuation in simultaneous manner [27]. The critical nature of scenarios must be considered to make the choices. Some of the approaches like Monte Carlo simulation and dynamic p-robust approaches are used to find the different scenarios and formulate the new paths for safety reliefs. Identifying various parameter values will be useful for robust reliable optimization. The rapid support decisions improve the response and recovery.

During emergency period the goal should be supplying the emergency services to the people with the goal of transport unit loss using the optimization model in the crowded conditions [28]. Bayes model theory with some optimization algorithms such as ACO for finding the efficient path from one place to another is needed to manage the congested conditions.

The remaining sections in this book chapter are organized in the following way. Section 4.2 describes the different methods/technologies related to predictive maintenance and safety management. Section 4.3 explains the issues of research work in this domain. Section 4.4 explains the new ideas found in this domain, and final section concludes and directions of research in future.

4.2.1 Survey on Predictive Analysis in Natural Disasters

Hasegawa et al. [28] introduced a device called as QUEST (Q-shu University experiment with steady-state spherical tokamak) to originate the large current of 50 KA. To minimize the operational costs and reducing the danger of unexpected failures, the contact resilience of multiple ohms was evaluated in order to make the predictive maintenance. Predictive maintenance was achieved, by monitoring the occurrences of problems, the failure was blocked and maintenance was executed as a planned manner. The alarm system was built and the experimental information was shared among the systems by building a network. For few devices, static Internet Protocol (IP) was assigned. For control devices, it had static IP and Media Access Control (MAC) addresses to evade the wasting of resources of network.

W. Gao [29] used ANN to predict the landslide using Grey System. Monotonously increasing characteristics and displacement of landslide need to be analyzed the time series data in the specific area where landslide happened frequently. Grey system combined with Evolutionary Neural Network (ENN) was proposed for prediction system. Composition of displacement and time series trend were used by ENN architecture. Backpropagation algorithm (BP) was used in ENN and applied in landslide Xintan. The results proved that this method was better in landslide prediction.

Matthews [30] applied Decision Theory to analyze about how earthquake were predicted accurately and reliably. Even though several optimistic estimates are used to identify the key parameters, the reliable function for the action should be used. He identified the self-organizing behavior of earthquakes prediction was based on two key factors.

• (i) the base rate of severe-quakes on the time scale of precursor, and
• (ii) the qualified costs of disregarding perfect predictions and answering to false alarms.

Even though with common values of these factors, the resultant least accuracy taken by the likelihood ratio needed to meet the correct predictions of significant earthquakes is not perfectly achieved in state-of-the-art methodologies.

G. Molchan and Romashkova [31] worked in the earthquake prediction and the quality of space related to time by using two- dimensional error diagram (n, τ) in which n represented the fraction of “failures-to-predict” and τ represented local-rate of alarm (averaged) in the space. The earthquake events during 1985 to 2009 (which had higher magnitude between 8 and 8.5) were taken as data samples and the M8 algorithm was used for prediction. The upper estimates were taken to handle the uncertainty.

Takahashi et al. [32] developed a network in the ocean floor to predict the earth quakes and tsunamis in the Japan Trench. By deploying seismometers and pressure sensors, the magnitude above 8 were collected. Based on the sensor details, the arrival time, tsunami height, and duration were predicted. They observed that the prediction accuracy depends on the height of tsunami and target point. But using seismometers, real-time prediction was done with buoy system and proved that estimation of target area was strongly correlated with target point.

Voigt et al. [33] used capabilities of earth observation used in the national and international level services, relief efforts, and security issues considered during the situations of disasters using the multi-source satellite images. Image analysis algorithms were used to map the tasks of resource lifecycle events to management support. These algorithms were used for rapid mapping of disaster response in tsunami, landslide, and forest fires.

Nico et al. [34] in Canada and Rauste et al. [35] in Finland used the thermal images to map the areas into wild forest fires using Advanced Very High Resolution Radiometer (AVHRR). Raw thermal data from satellite was taken as input about the amount and the number of truly burning fires was predicted.

4.2.2 Survey on Safety Management and Recovery

Hanna et al. [36] discussed the various issues regarding Discovery Recovery Planning (DRP). This includes recovering the damaged computers, connected to the hospital groups. The planning assured about health-care applications to reach the people affected during the disaster without interruptions with the goal of hospital and nursing functions.

Alabdulwahab [37] implemented some technology-based support to business to recover from the disaster. The data, hardware, and software are the major assets to the IT companies. Different strategies to recover the data were framed out.

Sanderson et al. [38] examined the one of the essential needs called as transitional shelter that was used in Haiti, analyzed some of the decision-making process could be used for making the adoption policy and making T-shelters. Based on the decisions, the results were compared with different responses planned in urban areas, post-disaster recovery, and redevelopment skills.

Barnes et al. [39] developed a novel method for classification and detection of structures using high-resolution satellite images. The features of images in near shore area were used for response planning during emergency. Both the storm impacted and non-impacted features were extracted from the images especially in blocked routes and used to automatically finding the discovery and rescue areas. The σ-tree template structures were used to reduce the time needed for training and it did not use high amount of computing resources.

Harris and K. Laubsch [40] surveyed different kinds of the airships at different height level from the earth to capture the live videos to enable the agile response. Flight ceilings of airships are designed depending on the tallest buildings in disaster affected area.

Moore [41] identified the different issues by interactions with electrical workers, managers in corporate, and professionals in safety management in a national hospital facility. The challenges in the recovery processes were identified to explore the electrical safety and disaster preparedness. Few case studies were done and the following were identified.

• (i) Elaborated challenges when handling the recovery project afterward an electrical disaster what affects employers and pendant legal action.
• (ii) Energized work plan to be noticed by safety professionals and executives.
• (iii) After the hazardous event, maintenance and safety culture to be followed by employers.

Slinn [42] proposed recovery system for railway management. After the interlocking failure happened, the communication between the sensors and control point the pictures were shared. From that picture, the details of train, states were identified to decide about what trains should be removed. To get the reliable pictures in railway management, COT technology was used. To assure the display the remote fail-safe image, encryption was suggested. For enforcing image lifetime, embedded image read-back was used.

4.2.3 Survey on Optimizing Solutions in Natural Disasters

Shao [43] surveyed the potential threats to the IT-based companies and proposed a discrete optimization model to use the redundancy mechanism for critical functions in recovery planning phase of recovery management. The overall objective of the author was maximizing the survivability of the IT functionalities in the organization. Each function was allocated with different level of redundancy based on priority. Dynamic probability–based programming was applied. In case of no redundancy, the model reduced the reliability related problems and lead to fault-tolerance.

Lu et al. [44] suggested the optimization method in urban places for disaster prevention using Geographic Information System (GIS) technology. He worked for optimal (by fast estimation) fire extinguishing equipment in a power grid during fire-disasters. The quick calculation method was proposed for identifying risk index in the transmission line. Hunan provincial power-grid was taken as case study. By using the computing resources, risk index was calculated soon to deploy the fire extinguishing equipment. The time of calculation was less than five minutes because 30 trillion operations were performed per second.

Unmanned Aerial Vehicle (UAV) was used by Zhang and J. Liu [45] because of wider coverage and low-vast, fast deployment in post- disaster environments. Two cooperative UAVs were used: one for downlink transmission and another one on rescue vehicles for emergency response. The data were sent as packets transmitted from an UAV to a vehicle, if time duration for the truck enclosed, UAV is more than the stated average network access delay. Performance metrics given for the network and this method gave good numerical results. Optimal setting of network parameters achieved optimal performances in the post-disaster areas.

Cheng-fang Wang and Yi-min Sun [46] analyzed different types of disaster in urban areas. Several case studies were done to identify the basic correlation between the urban morphology optimization and corresponding disaster prevention. The parameters such as location of the urban place or city, environment, size of the urban, structure, and the direction of the development of urban were used as parameters to analyze. The spatial information from GIS was used to optimize the locality of urban toward disaster direction.

Yuan et al. [47] proposed Resilient Distribution Network Planning Problem (RDNP) for coordinating and distribution of resources. The objective was to minimize the damage of the system. Hurricanes create huge losses economically and casual life of humans. The problem was formulated as two stage strong optimization model. Uncertainty was solved by N-K contingency method. The system was validated using micro-grids and proved that resilience was good during natural disasters. Yuan et al. [48] worked to improve the grid resilience during hurricanes, and the consequences such as physical and cyber-attacks can be reduced by optimal plans. One of the optimal solutions is planning the decisions to coordinate the DG replacement, distribution of the power flow model, and hardening.

Changfeng et al. [49] developed optimization based on Bayes risk function for managing emergency resources in the transportation network using a case study to direct the composite association between supplies distribution and selection of path, using the factors such as travel time uncertainty and the lasting capacity of road during distribution. Emergency logistics structure of the network was taken and total loss of disaster area was predicted using Bayes risk function. ACO algorithm was used in crowded situations and proved that the optimization model behaved in better manner during the congested conditions.

4.3.1 Forward-Looking Strategic Vision (FVS)

The International Data Corporation (IDC) analysts identified that in most situations, most of the real-time environments are yet to deploy the sensors for data collection. So, many of the IoT applications are suffering to collect the real data. Based on the applications even though sensors are deployed and communications are done properly, huge data are collected, the users of IoT may not get satisfaction because of lacking data analytics skills and security concerns, because of not having or developing an FSV.

Many mathematical models of statistics and ML (sometimes combined as hybrid) approaches are needed with big data to get the scientific and quantifiable results. If suitable approaches are used with optimal parameters, then the goal can be attained. If the data set collected is biased, then for new kind of data, the result may be incorrect. The balanced interplay of insights of data analysis gives informative and successful solutions.

4.3.2 Availability of Data

The benefits of disaster management depend on the available and appropriate data. The data collected may not have the sufficient information, not in the structured format, and aggregated from different locations. Sometimes, the data may be irrelevant. False and incomplete data leads to uncertain outcomes. Sometimes, mathematical models have to be used for identifying missing and incorrect values in the collected data. Therefore, to make the data to be usable directly, some pre-processing methods are to be applied.

4.3.3 Load Balancing

There are four layers in IoT. They are sensor and actuator layer (sometimes, these two layers are combinedly called as perceptual layer), network layer, data processing, and application layers. Sensor and actuator layer have the direct access to gateway. In edge computing, most of the data processing work is done in the gateway. Most of the computing resources are needed at the local gateway.

In fog computing, data should be transferred to the data processing layer that is far away from the sensors. Therefore, based on the type of computing, allocation of resources is a challenging task. Dynamic scheduling model with priority based one should be selected, because of the nature of the application. Based on the scheduling method selected with real-time situation, the available resources should be balanced. Therefore, the resource allocation task makes the computing as distributed computing.

4.3.4 Energy Saving and Optimization

The overall energy consumption is calculated as the summation of autonomous consumptions created by machining, transfer of data at all the stages of hazardous events, and mining algorithms. In edge computing, much energy will be consumed at the local gateway and in the fog computing much energy will be consumed in devices at the data processing layer. Most of the real-time devices are battery based and rechargeable, so the energy-saving mechanism is to be considered. The different kinds of modes like sleep and wake up can be used and a particular event can be used for triggering sleep mode to active mode to save the energy.

Efficient allocation of computing resources is a challenging task because of

• (i) Processing power limitation,
• (ii) Rapid computationally intensive applications,
• (iii) Energy consumption analysis needs the complex system by consideration of spatial attributes.

Optimization algorithms can be utilized to identify the threshold level for each kind of sensor about, under what values the switch over between wake up and sleep mode should be done.

4.3.5 Cost Benefit Analysis

Sometimes, the real cost benefit may not be identifiable in accurate way. Few applications can be done only in simulation, but not in the real world. It may be because of the cost of real-time implementation, nature of the event. It should not be trial-and-error effort.

Before going to real implementation, simulations can be done. Based on the success of simulation prototypes can be created to analyze the real-time risks and improve the efficiency. The task of identifying idea, simulation, prototyping, and evaluation should be iterative tasks until all the stakeholders get satisfaction.

4.3.6 Misguidance of Analysis

Many attempts can be done to find the approaches for predicting the major earthquakes, and sometimes, they are misguided because of the dynamic nature of the event. This may be created because of false alarms, device hacking, inefficient communication, and coordination lacking. The output of the ML algorithms depend on the based on the data samples fed into the algorithm. If some data collected has incorrect values, then misguidance may happen.

4.4.1 Data Driven Reasoning

In many real-time problems, it is possible to find all the possible solutions beforehand and select the suitable one. Safety and recovery during disaster is such a kind of problem. But in the data driven approaches, multiple data sets in history can be collected, and experiences of them can be utilized to provide the solution in the basis of evidence-based approach. Problem solution can be created by rule based that gives solution similar to real prediction. Using advanced technologies for data collection and analysis techniques (with reasoning), effective recovery and reconstruction plans can be done and executed.

4.4.2 Cognitive Ability

Cognitive ability is a general rational ability that involves reasoning, solving problems, preparation, abstract thinking, comprehension of complex ideas, and self-learning using the experience. It needs more attention. Even though enough guidance were given to the people, the police, emergency handlers, NGOs, and others involving in the reconstruction may experience cognitive disruption and this can affect their response in rescue-related tasks. This may happen because of the unexpected scenarios faced by all the responding people. Based on the cognitive skills trained in the system, trained system (or algorithm) should make clear decisions to guide in stressful situations.

4.4.3 Edge Intelligence

Using edge intelligence, the sensing device takes the control of its data and communications organization, improves the accuracy and quality, and reduces the time delays. Intelligent edge algo enlarges set of associated systems closer to the users. The intelligent algorithms are used either in the device or in the edge of a network with particular coverage of locations should be used for more critical applications. Safety applications in disaster response got real-time experiences and insights, deliverable in a context aware manner.

4.4.4 Effect of ML Algorithms and Optimization

ML algorithms have better attainments than human beings, using the algorithms like regression and classification. Using optimization methods, the minimal resources would be utilized or minimal time can be taken and the better solutions can be found. Therefore, the overall execution cost can be less with user defined time. There may be multiple feasible plans for response and recovery generated. But based on the real-time scenario, optimal plan must be selected by using meta-heuristic optimization algorithms.

4.4.5 Security

The data collected and the decision made under different scenarios should be shared within the people or authorities who involved in the safety management application. The security for data, network, and data storage devices must be provided to avoid the active and passive attacks which are planned by the hackers. The risk should be predicted in earlier and standard encryption and corresponding decryption algorithms can be used to enhance the security. Based on the use cases, the appropriate algorithms can be utilized.

4.5.1 Conclusion

Disaster management is as the group of activities, management of resources, and tasks for managing with all humanitarian features of emergencies, preparedness, response, and also recovery in direction to reduce the impact of disasters. Proper analysis and prediction on a time scale of decades makes the planning and response arrangements. During the planning and response, the factors such as safety of the people, safety of physical assets, and providing food, medicine, and temporary shelters must be taken care.

Based on the experience, the construction ideas, development in urban areas, green spaces, and the municipal affairs provide multi-discipline power to make the safe city. Different countries have different kinds of natural disasters based on the region in terms of latitudes, climate changes, industrialization and other conditions of nature.

4.5.2 Future Research

For further research, the assets can be categorized into essential, shelters, companies’ infrastructure, communication lines, vehicle track, road, and standing water, train track, based on that priority can be applied. Based on the category, appropriate plan can be applied to reduce the damage of assets. Different security levels can be defined based on the different use cases. Based on the use case types, recovery plans can be generated.