smartresilience - d4.2 resilience indicators for scis ... · assessing resilience of scis (#2) by...

PUBLIC DELIVERABLE

H2020 Project: Smart Resilience Indicators for Smart Critical Infrastructure

D4.2 - Resilience indicators for SCIs based on big and open data

Coordinator: Aleksandar Jovanovic EU-VRi Project Manager: Bastien Caillard EU-VRi

European Virtual Institute for Integrated Risk Management Haus der Wirtschaft, Willi-Bleicher-Straße 19, 70174 Stuttgart

Contact: [email protected]

SMART RESILIENCE INDICATORS FOR SMART CRITICAL INFRASTRUCTURES

© 2016-2019 This document and its content are the property of the SmartResilience Consortium. All rights relevant to this document are determined by the applicable laws. Access to this document does not grant any right or license on the document or its contents. This document or its contents are not to be used or treated in any manner inconsistent with the rights or interests of the SmartResilience Consortium or the Partners detriment and are not to be disclosed externally without prior written consent from the SmartResilience Partners. Each SmartResilience Partner may use this document in conformity with the SmartResilience Consortium Grant Agreement provisions. The research leading to these results has received funding from the European Union’s Horizon 2020 Research and Innovation Programme, under the Grant Agreement No 700621.

The views and opinions in this document are solely those of the authors and contributors, not those of the European Commission.

Resilience indicators for SCIs based on big and open data

Report Title: Resilience indicators for SCIs based on big and open data

Author(s): P. Klimek, R. Lo Sardo, V. Maraglino, A. Choudhary, T. Knape, A. Jovanović

Responsible Project Partner: MUW

Contributing Project Partners: R-TECH, EU-VRI, SwissRe, AIA

Document data:

File name / Release: SmartResilience_D4.2-BigDataRI_v07bc26042018 Release No.: 2

Pages: 42 No. of annexes: 1

Status: FInal Dissemination level: Public

Project title: SmartResilience: Smart Resilience Indicators for Smart Critical Infrastructures

Grant Agreement No.: 700621

Project No.: 12135

WP title: Defining classic and deriving smart Resilience Indicators (RIs)

Deliverable No: D4.2

Date: Due date: April 30, 2018 Submission date: April 30, 2018

Keywords: big data, network science, data mining, infrastructure networks

Reviewed by: I. Shapira Review date: April 20, 2018

M.-V. Florin Review date: April 20, 2018

Approved by Coordinator: A. Jovanovic Approval date: April 26, 2018

Vienna, April 2018

SmartResilience: Indicators for Smart Critical Infrastructures

page i Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Release History

Release No.

Date Change

1 April 9, 2018 Preliminary version

2 April 26, 2018 Final version


page ii Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Project Contact

EU-VRi – European Virtual Institute for Integrated Risk Management Haus der Wirtschaft, Willi-Bleicher-Straße 19, 70174 Stuttgart, Germany Visiting/Mailing address: Lange Str. 54, 70174 Stuttgart, Germany Tel: +49 711 410041 27, Fax: +49 711 410041 24 – www.eu-vri.eu – [email protected] Registered in Stuttgart, Germany under HRA 720578

SmartResilience Project

Modern critical infrastructures are becoming increasingly smarter (e.g. the smart cities). Making the infrastructures smarter usually means making them smarter in the normal operation and use: more adaptive, more intelligent etc. But will these smart critical infrastructures (SCIs) behave smartly and be smartly resilient also when exposed to extreme threats, such as extreme weather disasters or terrorist attacks? If making existing infrastructure smarter is achieved by making it more complex, would it also make it more vulnerable? Would this affect resilience of an SCI as its ability to anticipate, prepare for, adapt and withstand, respond to, and recover? What are the resilience indicators (RIs) which one has to look at?

These are the main questions tackled by SmartResilience project.

The project envisages answering the above questions in several steps (#1) By identifying existing indicators suitable for assessing resilience of SCIs (#2) By identifying new smart resilience indicators including those from Big Data (#3) By developing, a new advanced resilience assessment methodology based on smart RIs and the resilience indicators cube, including the resilience matrix (#4) By developing the interactive SCI Dashboard tool (#5) By applying the methodology/tools in 8 case studies, integrated under one virtual, smart-city-like, European case study. The SCIs considered (in 8 European countries!) deal with energy, transportation, health, and water.

This approach will allow benchmarking the best-practice solutions and identifying the early warnings, improving resilience of SCIs against new threats and cascading and ripple effects. The benefits/savings to be achieved by the project will be assessed by the reinsurance company participant. The consortium involves seven leading end-users/industries in the area, seven leading research organizations, supported by academia and lead by a dedicated European organization. External world leading resilience experts will be included in the Advisory Board.


page iii Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Executive Summary

Modern critical infrastructures—in particular, smart critical infrastructures—routinely produce abundances of data, ranging from billing data over server log files to sensor data. The central idea in task 4.2 of the SmartResilience project was to identify such datasets for individual case studies and to investigate if and how such data, enriched with other open data sources, can be re-used to derive resilience indicators. A central challenge is that such data is typically collected with other purposes in mind than resilience assessments. It is therefore not at all clear how such data can be transformed into actionable knowledge by the means of indicators. To meet this challenge, it is necessary to go beyond state-of-the-art approaches in dealing with such complex and dynamic datasets in order to prune out relevant information.

It is not feasible to develop or create a unified methodology enabling the derivation of all potential indicators from data from all possible infrastructures. However, some commonalities across different case studies could be established within the project. This allowed us to formulate a common framework that could be applied quite broadly—ranging from indicators for systemic risk in healthcare over organizational coordination in emergency response to early warning signals in industrial production systems. The key insight in these efforts was that most of the considered datasets can be conveniently represented as dynamic networks. By using appropriate techniques from network science and related fields, it is possible to identify elements in the data (specific sensors, emergency responders, healthcare provider, …) that are critical in determining the functionality or resilience of the described infrastructure, for instance, because their removal could impair the functioning of the entire infrastructure system. Network measures that are sensitive to such properties can be used as the basis to formulate resilience indicators. In this report we give four in-depth examples for such a derivation of indicators, with application areas that include healthcare systems, emergency response, transportation systems, and a petrochemical plant as example for an industrial production system.

Big-data based indicators are also particularly useful to address the issue of (inter-)dependences between different infrastructures. Here the work in T4.2 overlaps with the work reported in T2.3, to which we refer the interested reader. In this report we want to stress that the data generated within the SmartResilience project itself, namely the structured collection of dynamic checklists of issues and indicators for particular infrastructures and threats, can also be leveraged using machine learning approaches. Therefore, we build a recommendation system based on collaborative filtering, that allows to some extent to automatically formulate assessment checklists for “infrastructures of infrastructures”. In order to do so, it is required that the project partners and case study leaders adhere to the use of recommended and core issues and indicators in their assessments as far as possible.

Finally, we recognize the fact that not each end user of the project has the resources to implement data-intensive and methodologically demanding procedures in his or her assessments. We therefore developed a tool—fully integrated into the SCI dashboard—that allows end users to leverage the work described here without the need to carry out the mathematical computations involved in deriving the indicators themselves. The workflow is summarized in Figure 1 and consists of the following steps. The end user provides input in the form of a dynamic checklist. He or she can then select indicators that should be (but need not be) assigned values that result from data analysis. The user can then download a template for the specific indicator with instructions on how to insert data from a certain infrastructure. This template is then uploaded to a data analytics background service that is available through the SCI dashboard. All that is then needed to derive the indicators is to click on a button to perform the analysis. The output is displayed in the form of assessments or benchmarks.


page iv Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Figure 1: SmartResilience use case workflow. The end user provides a dynamic checklist that

includes big data based indicators and uploads his or her data using a template. The SCI dashboard offers access to web-based deployments of the big data analytics services described in this report, which occur in the background. The output for the end user is an assessment that includes data-intensive and conventional indicators.


page v Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Table of Contents

1.1 Task 4.2 in the context of the SmartResilience project .......................................5 1.2 Overview of the methodology ..............................................................................5 1.3 Analysing dynamic, heterogeneous, interconnected, and multirelational

data.........................................................................................................................6 1.4 From data to indicators .........................................................................................9

2.1 Example I: CHARLIE ............................................................................................. 11 2.2 Example II: DELTA ............................................................................................... 15 2.3 Example III: ECHO ............................................................................................... 19 2.4 Example IV: GOLF ............................................................................................... 22 2.5 Indicators for other case studies ....................................................................... 22


page vi Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

List of Tables

Table 1: Big data vignette for CHARLIE. ............................................................................. 11 Table 2: Big data vignette for DELTA. ................................................................................ 15 Table 3: Big data vignette for ECHO. ................................................................................. 19 Table 4: Big data vignette for GOLF. .................................................................................. 22 Table 5: Selection of big and open data related indicators across different case

studies. .................................................................................................................. 23 Table 6: Big data vignette for interdependences. ............................................................ 25


page i Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

List of Figures

Figure 1: SmartResilience use case workflow. The end user provides a dynamic checklist that includes big data based indicators and uploads his or her data using a template. The SCI dashboard offers access to web-based deployments of the big data analytics services described in this report, which occur in the background. The output for the end user is an assessment that includes data-intensive and conventional indicators..........................................................iv

Figure 2: Illustration of the derivation of RIs for cascading failures using systemic risk models for the example of financial institutions. (A) A simplified version of the balance sheet of a bank is shown with different types of liabilities (red) and assets (blue) (figure taken from [13]). An external shock reduces the asset size of the bank. If the shock is higher than the net worth, the bank defaults. Other banks need to write off their loans to the defaulting bank—the shock propagates. How likely such a propagation is, depends on the network structure of financial relations. Overly simplified examples are shown for networks with (B) high and (C) low systemic risk. .............................. 7

Figure 3: Illustration of correlation network analysis to extract functional networks. (A) We consider a data set consisting of six different time series, which could be repeated measurements from sensors at a petrochemical plant. How a time series is related to the other one can be computed using appropriate similarity measures, such as (B) Pearson’s correlation coefficient, giving a matrix in which cells indicate how likely a change in one sensor coincides with the same (or opposite) change in another sensor. (C) After appropriate filtering, this correlation matrix can be represented as a functional network that represents the key dependences in the original dataset. This network can now serve as a basis to identify nodes with a specific role, e.g., in the transmission of failures. ......................................................................................... 8

Figure 4: In the case studies that are covered in-depth in this report, conventional issues and indicators could be mapped to network measures, that in turn were evaluated by extracting the relevant networks from the original data. ............. 9

Figure 5: Illustration of different centrality measures [12]. The white nodes are good spreaders in terms of how fast other nodes can be reached from them—this is captured by centrality measures like the Katz prestige. Each shortest path (dashed line) from nodes on the left hand side to the right hand side has to run through the dark-grey node, it has a high betweenness centrality. .............................................................................................................. 10

Figure 6: All healthcare providers are shown as circles on a map of Austria (top right), with a zoom into a region of Vorarlberg (top left). In his zoomed view, the color of the healthcare provider indicates their type (ochre for primary care, turquoise for secondary care or specialists, and blue for hospitals) and the size corresponds to their number of patients treated at a particular day. Note that the positions of the providers have been randomized within a district for reasons of data privacy. It is then possible to follow the movement of individual patients within this map between different healthcare provider (bottom). ............................................................................. 12

Figure 7: Modelling inaccessibility or inoperability of healthcare providers. Left: for each provider the number of patients is known (here a doctor with 7137 patients


page ii Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

is shown) and how any of those patients have also been treated elsewhere (links to other doctors). Right: Should the original doctor be removed, his or her patients are distributed according to those link weights (yellow links). As some providers might now exceed their capacity, a secondary cascade of displacements is triggered (blue lines). The degree of red of the circles shows how close the individual providers are to their capacity. ....................... 13

Figure 8: Isochrones are useful to determine the reachability of care providers, such as hospitals. To the left we show a map of Vienna with markers for places where a hospital is located (here the General Hospital of Vienna is highlighted). To the right we show isochrones for this hospital, shaded areas from which the hospital can be reached in the same time. .................... 14

Figure 9: The models have been deployed using an short interactive throw projector, which allows to define and play through scenarios by touching the model, as was done here for a demonstration of the technology in front of the Smart City Council of the City of Vienna. ............................................................ 15

Figure 10: Translating communication network properties into indicators. The clustering coefficient, betweenness centrality, and Katz prestige can be related to the presence of redundant, vulnerable, and efficient information flows, respectively. .......................................................................................................... 16

Figure 11: Communication networks for all messages sent on day 1 (left) and for day 2 (right). Each node corresponds to one organization, directed links between two nodes indicate information flows. Node colors correspond to hierarchical levels, i.e., blue for the high (national or regional) and green for low (local) levels. Node size is proportional to the number of sent messages. Both networks consisted of multiple communities of nodes. However, on day 2 there was a much higher number of information flows between nodes from different communities all across the network as compared to day 1 ............................................................................................... 17

Figure 12: 3d scatter plot of the network indicators that represent efficiency, redundancy, and vulnerability. Each point corresponds to the communication network at a given time with participants being either prepared (red circles) or unprepared (blue squares). There is a clear tendency that prepared participants showed higher efficiency, lower vulnerability and less redundancy in their communication. .............................. 18

Figure 13: Communication networks extracted on site with live data during the 2017 emergency drill. The node colors indicate organizations with high vulnerability (left) and efficiency (right), respectively. This information was used during the de-briefing of the exercise. ....................................................... 19

Figure 14: Overview of workflow of the big-data indicators in ECHO. Based on the sensor data NIS Pančevo Oil Refinery, a statistical time series analysis was performed to detect outlier measurements. Whether such outliers signal that other components might be affected as well is determined in a correlation network analysis that informs early warning indicators. These indicators can be used in DCLs for short-term monitoring. ............................... 20

Figure 15: Deriving RIs based on sensor time series. Left: anomalies in the data of individual sensors can be detected by measuring by how many standard deviations the current measurement exceeds the values normally observed. Outliers are shown as red spikes. Right: Whether such an outlier is a precursor that further deviations should be expected or not can be determined from the functional network of the sensors. Here, each node represents a sensor and two sensors are linked if the measurements at these sensors typically correlate. Early warnings are those outliers that occur at nodes with a high centrality in the functional network (shown as the degree of red of the node colors). Note that the functional network has well discernible modules (two of them being highlighted by green circles) that closely mimic the actual physical layout of the plant. ................................ 20


page iii Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Figure 16: Early warning indicator for the event observed in the data. The indicator for the sensor shown here shows no signals until briefly before the end of the observation window where the output quality deteriorated, where a strong signal is picked up. This event was indeed the strongest observed event in the data, as shown by the ranks of detected early warnings. ........................... 21

Figure 17: Overview of the indicator-based approach to interdependences. The case-study infrastructures are shown as large green circles, indicators as small blue circles. Links connect infrastructures and indicators if the indicator has been used in the assessment of an infrastructure. ............................................ 26

Figure 18: SmartResilience use case workflow. The end user provides specifications of a DCL as input and uploads his or her data. The SCI dashboard offers access to web-based deployments of the big data analytics services described in this report, which occur in the background. The output for the end user is an assessment that includes data-intensive and conventional indicators. ....... 29

Figure 19: Illustration of workflow in the SCI dashboard to utilize the big data indicators. After the DCL has been generated, existing big data indicators can be selected. A template is then offered that allows the user to upload his or her data and run the analysis with the click of a button. In the case shown here, the result is an assessment of the resilience level in the ECHO case study...................................................................................................................... 30


page iv Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

List of Acronyms

Acronym Definition ABM Agent-based model

BLEVE Boiling liquid expanding vapor explosion

CI Critical infrastructure

DCL Dynamic checklist

RI Resilience indicator

SCI Smart critical infrastructure

SME Small and medium-sized enterprises


page 5 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Introduction—From Big Data to Indicators

1.1 Task 4.2 in the context of the SmartResilience project Today’s infrastructures routinely produce abundances of data in various forms, be it log files, billing data, traces of online activities, historical sensor data, etc. This routine data is typically collected for purposes that have nothing to do with understanding the resilience of the system that generated it. Nevertheless, with the increasing availability of such data and ever-growing computational resources, it is becoming increasingly clear that such routine data captures important aspects of the functionality and/or resilience of the infrastructure that generated the data, be it under normal or abnormal conditions. This data is too expensive to waste. The central idea in the project was therefore to explore the re-use of such routine data—leveraged by other open data sources—in the formulation of resilience indicators. The aim of T4.2 in the project was to show how this can be done using a case-study lens, with a focus on those cases in which particularly useful datasets could be identified (see D5.1 [1]).

The central challenge in this task is posed by the fact that the raw data is usually high-dimensional (many different variables), complex (these variables can be closely related with each other or completely independent, or anything in-between) and dynamically changing over time. It is therefore not at all clear if and how such massive datasets can be collapsed onto one or several numbers that help us to characterize the resilience of an infrastructure—numbers that can be used as resilience indicators (RIs). This is the classic challenge in data analytics. Here we will show how a combination of different number-crunching techniques (network science, machine learning, statistical hypothesis testing, agent-based models, regression analysis, GIS-based methods, see sections 2 and 3) was successfully applied in different case studies to transform complex routine data into actionable knowledge in the form of RIs. In particular, the methods developed here show how data from case study owners in combination with data from the public domain can be used to define indicators for specific issues in a resilience assessment. These indicators can (but need not!) be “filled” and/or updated with values extracted from the data in an automated way—in this sense these RIs are “automated” or “autonomous”. The use of such indicators in the project is completely identical to the use of more conventional or “supervised” RIs, such as those defined in T4.1. The “big data indicators” can be used in assessments by means of dynamic checklists (DCLs) and are fully compliant with the indicator-based SmartResilience methodology developed throughout WP3. Wherever possible, they have been integrated into the SCI dashboard, where they can be used in assessments to be made throughout WP5.

1.2 Overview of the methodology There is no such thing as one general and unified quantitative methodology to derive all RIs that could ever be of interest for all existing infrastructure in a way that would be applicable to each and every data source. However, there are some shared characteristics between many infrastructures that allow one to formulate a structured approach that was successfully applied within the context of several case studies in the project. The fact that modern infrastructures are becoming increasingly smart and interconnected means that they can only be properly characterized by multiple interdependent networks whose functionality levels depend on one another [2],[3]. This is true for infrastructures of infrastructures (as addressed in INDIA) and for infrastructures themselves that can often be represented by specific types of networks (e.g., multi-modal transportation networks, organizational and institutional networks that operate an infrastructure, etc.). Failure of one component of the system (say, a certain sensor in a refinery) may lead to a failure of another


page 6 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

component (e.g., mounting pressure in a vessel) which might in turn trigger other failures that eventually lead to a breakdown of the entire system (a boiling liquid expanding vapor explosion, BLEVE) [4]. To understand the likelihood, impact, and consequences of such events it is necessary to identify chokepoints, bottlenecks, redundancies, as well as the spreading behavior within such networks [5].More sophisticated approaches to measure RIs are particularly useful whenever the data from the considered system has one or several of the following characteristics:

(i) Dynamic (the data that characterizes the system changes frequently in time) (ii) Heterogeneous (many different types of elements define the functionality level of the system) (iii) Interconnected (functionality of the system is contingent on a large set of connections between

potentially heterogeneous elements) (iv) Multirelational (connections can be heterogeneous themselves, including linear, non-linear,

threshold models, or other types of dependencies)

To set the stage, we will review some seminal contributions on how to derive RIs for infrastructures that are described by dynamic, heterogeneous, interconnected, and/or multirelational data. These contributions informed the choice of the methodology that we then applied in individual case studies, as discussed in section 2. Also, note that these methods where complemented with other standard data analysis methods, including time series analysis (for instance, sensor time series in ECHO), regression models (e.g., indicator validation in DELTA), GIS-based analyses (accessibility in CHARLIE and transport times in GOLF) and other machine learning approaches (e.g., the collaborative filtering method described in section 3) whenever necessary. These methods have been well described and established in past decades, see for instance [6],[7],[8] for further literature. Until recently, however, there existed a substantial methodological gap when it comes to working with the types of dynamic, heterogeneous, interconnected and multirelational data that one typically encounters when working with infrastructure systems [2],[3]. The main work effort in this task therefore consisted in finding ways to bridge this gap, which is therefore also the main focus of this report.

1.3 Analysing dynamic, heterogeneous, interconnected, and multirelational data A central challenge that is posed by the presence of interconnections in and between critical infrastructures (CIs) is that the likelihood of a failure of a certain component (or the impact of an adverse event) is no longer a property of the component itself, but rather depends on all other components on which the original component depends, and the components on which those depend, and so on [9], [10]. In such cases it becomes necessary to consider network properties of the system. In brief, the main idea is to consider the hypothetical removal of one of the nodes in the network (e.g., a power node in an electrical grid where some initial failure occurs). Owing to the dependences encoded in the networks that represent the infrastructures, each node connected to the initially removed node might fail as a consequence. In this way a localized failure “spreads” on the network, similar as the spread of diseases on contact networks. The central result of such analyses is that given a particular initial localized failure, the way in which this failure affects the remaining parts of the network depends solely on certain structural characteristics of the network (often in a highly non-trivial and non-linear way), such as the number of nodes that can be reached within a given time or number of steps on the network [11]. That is, the behavior of the system can be understood on the basis of the topology of the underlying network [8],[12]. These topological characteristics can then serve as RIs. In the case of the cascading failures on power grids, for instance, it was found that the robustness of the system can be understood based on the so-called percolation behavior of the networks—loosely speaking the likelihood that there exist directed paths to walk from one node to any other node along links in the network. For readers that are not familiar with basic notions to describe networks, we give a brief overview of the essential concepts of network science in Box 1. From the methodological point of view, we followed a similar strategy when deriving RIs based on communication networks in the case study DELTA where we consider different emergency responders that depend on each other through their need to exchange information. Vulnerability can then be assessed by identifying those actors whose removal would most impede the flow of information on the network (rather than the flow of electrical power as in the case of the blackout).


page 7 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

The risk that an entire network will fail due to an initially localized failure is also called systemic risk [13]. The need to address risk at the systemic level has been reinforced in the wake of the global financial crisis that started in 2007 [14],[15],[16]. In this context, financial institutions can be represented as nodes in a network with links that represent lending relations. Consider the case where a bank A borrows money from bank B and assume that bank A is hit by an external shock (say, loss of assets due to an extreme weather event). If the shock is larger than its capital reserves, bank A fails, see Figure 2A. Now bank B has to write off the loan it gave to A. If this loan was larger than B’s capital reserves, the shock starts to propagate on the network, eventually also affecting other banks that had no direct relation with A whatsoever. In this case, links do not encode linear dependencies but rather threshold models: a shock is only transmitted along a link if it has a certain effect on the bank’s capital. A critical amount of evidence that suggests that such kind of network mechanisms are what drove world economies from the initial failure of single financial institutions (e.g., Lehman Brothers) to the largest recessionary shock recorded in human history [13],[14]. It turns out that the likelihood of such cascading defaults—systemic risk—depends strongly on the topology of the network of financial relations [15],[16]. Figure 2B shows an example of a financial network where the institutions are hierarchically ordered (shown as different shades of red). If a node upstream is hit by a shock, all nodes that sit downstream are at risk. Red nodes have a high systemic risk. The opposite is the case in Figure 2C, where each pair of institutions are mutually linked and the impact of an initial shock can be absorbed and spread out across several different nodes; systemic risk is low. This shows how the behavior of the system with respect to adverse events can again be understood based on the topology of the underlying networks, in this case by means of so-called centrality measures (see Box 1). We pursued a similar approach in case study CHARLIE, where we consider networks of healthcare provider with links that correspond to the exchange of patients. Surges in the number of patients to be treated (due to, e.g., disasters, removals of a single doctor or, say, inaccessibility of a certain hospital) might trigger cascades of patient displacements (patients seeking or requiring to be transported to different points of care), resulting in an over-crowding of other healthcare provider. In essence, this is again a problem of modelling systemic risk on networks.

Figure 2: Illustration of the derivation of RIs for cascading failures using systemic risk models for the example

of financial institutions. (A) A simplified version of the balance sheet of a bank is shown with different types of liabilities (red) and assets (blue) (figure taken from [13]). An external shock reduces the asset size of the bank. If the shock is higher than the net worth, the bank defaults. Other banks need to write off their loans to the defaulting bank—the shock propagates. How likely such a propagation is, depends on the network structure of financial relations. Overly simplified examples are shown for networks with (B) high and (C) low systemic risk.

In the examples we discussed so far we focused on actual, directly observable networks, e.g., power grid networks. However, in data analysis it is often the case that the “real” networks cannot be directly observed but reveal themselves through functional dependences between different observable variables. Such kind of functional networks can be extracted with a technique called correlation network analysis [12],[17]. To


page 8 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

illustrate the principle, consider the example where you have time series data from six different sources (which could be sensors in a petrochemical plant, as was done in the case study ECHO, but also stock prices, gene expressions, etc.), see Figure 3A. The idea is to define a similarity measure for each pair of time series, such as their Pearson’s correlation coefficient [18], see Figure 3B. This coefficient gives an estimate for how likely a change in a time series (say, increase in pressure in a specific pipe) coincides with the same (or opposite) change of another time series (such as the increase in temperature in another vessel). Using appropriate filtering techniques (such as taking the strongest correlated links, or some more sophisticated approaches like multiple hypothesis testing corrections [19] and network backboning [20]), it is possible to extract a functional network from the correlation matrix, see Figure 3C. The resulting functional network can now be analyzed using standard network measures to identify hubs, good spreaders, bottlenecks, etc., see Box 1. There is no guarantee that a link in the functional network represents a direct causal relation, as there could be a third, non-observable variable that drives the observed correlation. Nevertheless, the topology of the functional network often allows one to understand which nodes are the most likely candidates for explaining specific functional properties of the network, such as systemic risk, the likelihood of cascading failures, or other type of RIs.

Figure 3: Illustration of correlation network analysis to extract functional networks. (A) We consider a data

set consisting of six different time series, which could be repeated measurements from sensors at a petrochemical plant. How a time series is related to the other one can be computed using appropriate similarity measures, such as (B) Pearson’s correlation coefficient, giving a matrix in which cells indicate how likely a change in one sensor coincides with the same (or opposite) change in another sensor. (C) After appropriate filtering, this correlation matrix can be represented as a functional network that represents the key dependences in the original dataset. This network can now serve as a basis to identify nodes with a specific role, e.g., in the transmission of failures.

Note that while correlations (as described above) between two elements typically capture linear dependence, the resulting functional networks can show highly nonlinear behavior on a systemic level, a property that is sometimes referred to as “emergence” [8],[12]. For instance, consider the spreading of an infectious disease on a social network [21]. The more time you spend in the proximity of an infected person, the higher is the chance that you get infected too (linear dependence). However, the property that a particular disease will become endemic and spread on such a network is highly non-trivial. It can be shown that as long as the numbers or durations of contacts are small enough, the disease cannot spread on the network (infections will remain local and eventually die out). However, for many types of networks there exists a critical epidemic threshold in the number of connections beyond which even a single infected individual can spread the disease to each other individual [21]. It can be shown that this is exactly then the case when there exists a path in the network from each individual to each other ones and the time which it takes to traverse such a path is smaller than the recovery time of infected individuals. These properties make networks particularly suited to study the emergence of collective critical phenomena, such as tipping points in ecological systems, systemic risk in financial markets, or the epidemic spread of diseases.


page 9 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

1.4 From data to indicators Every data that can be stored in a relational database (basically everything that is stored as one or several interconnected tables) is some type of network (e.g., a table that links specific purchases to customers, or power generating stations to high voltage transmission lines that connect them) [22]. If this data is “simple enough”—when relations do not change over time or all changes that occur are independent from each other—the network view can be safely neglected. However, in all other cases and, in particular, when relations in the data are dynamic, heterogeneous, interconnected, and multirelational, the behavior of the underlying system can only be understood on a quantitative basis by explicitly modeling the dependence structure within the data. The concrete way in which this kind of modeling should be approached depends on the concrete context of the scenario to be considered. However, as a general guideline the following can be said.

1. Identify which aspects of the functionality of an infrastructure are captured by the data, see also D5.1 [1]. This informs the applicability and limitations of all subsequent results.

2. Extract the key relations from the data by mapping it on networks. For “actual”, physical network data this process is straight-forward, for data mining purposes beyond that the use of functional networks is recommended.

3. From those networks, indicators for the aspects of the functionality level captured by the network can be modelled using network measures, see Box 1 or [8],[12]. For instance, if one is interested in the nodes with the strongest relations to other nodes, the node degree might be a useful indicator. To identify nodes that are important in maintaining relations or flows between different parts of the network, one should consider betweenness centrality. To identify nodes that are highly accessible from all other nodes, closeness centrality is a suitable indicator. Nodes that are good spreaders, i.e., that can affect many other nodes in a short amount of time, are best identified using Katz prestige or eigenvector centrality, see also Figure 4.

Note that in some cases steps 2 and 3 can take on a rather trivial form, in particular in cases where the indicators can be read off directly from the data, e.g., in the form of raw counts of certain types of events. The technical implementation of the process described here is elaborated in greater detail in section 4.

Figure 4: In the case studies that are covered in-depth in this report, conventional issues and indicators could

be mapped to network measures, that in turn were evaluated by extracting the relevant networks from the original data.

Network measuresDegreeClustering CoefficientBetweenness CentralityKatz PrestigeCloseness CentralityDiameterTransitive TriplesNearest-neighbor degreeCharacteristic DistancePageRankPrincipal Eigenvalue...

Conventional Issue / IndicatorHow efficient is information transmitted through the system?How vulnerable is the system to random attacks?How susceptible is the system to targetedattacks?How many redundant motifs are in the system?Where do chokepoints exist in the system?Do people conform to the agreed-upon procedures?Are our procedures appropriate?

...


page 10 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Box 1: A network primer Networks consist of nodes that are connected by links [8],[12]. Links are either directed or undirected. Undirected links mean that two nodes mutually share a relation, whereas directed links typically represent an effect from one node to the other (e.g., a flow of power, people, or money). Mathematically, networks can be conveniently represented by so-called adjacency matrices, A. Each row/column in this matrix corresponds to one node. If there is a link from node n to node m, the matrix element Anm=1, otherwise it is zero. In undirected networks, Anm=1 always implies Amn=1. Two nodes that share an undirected link are called neighbors.

The most basic property of a node is its degree. The degree is the number of links attached to a node. For directed networks, one distinguishes in-degree (number of incoming links) and out-degree (number of outgoing links); for undirected networks those two are identical. The distance between two nodes in a network is the smallest number of links that must be traversed to move from one node to another on the network (link directions matter), which is the length of the shortest path between two nodes. Another important node property is the clustering coefficient, which is the probability that any two neighbors of a node are also neighbors of each other.

Many important structural properties of networks can be expressed by means of centrality measures for nodes, see also Figure 5. There are two classes of them: distance-based and random-walk based centrality measures. The most frequently encountered distance-based measures are closeness and betweenness centrality [23]. Closeness centrality measures the accessibility of a node in the network and is defined as the inverse average distance to each of the other nodes. Betweeness centrality measures how important a node is in terms of connecting to other nodes. Consider the set of all shortest paths between all pair of nodes in the network. A node’s betweenness centrality is the fraction of these shortest paths that involve the considered node.

The idea behind random-walk based centrality measures is to define importance of a node not based on how many links a node has, but on how important those nodes in turn are. In its simplest formulation, each node is as important as the other. In this case, centrality is simply given by the degree itself, the so-called degree centrality. In other cases, it might be more reasonable to assume that links with high-degree nodes are more important than links with low-degree nodes, which is described by eigenvector centrality [24] and Katz prestige [25] (the difference being that Katz prestige is also suitable to describe nodes with zero in-degree, as opposed to eigenvector centrality). Finally, Google’s PageRank algorithm additionally assumes that the importance of each node is split up across all of its outgoing links (that is, the higher the out-degree of a node, the smaller its centrality contributions to all nodes that have a link to it—your site is not as important as Wikipedia simply because it receives a link from there, but rather Wikipedia’s importance “splits up” across all of its links) [26].

Figure 5: Illustration of different centrality measures [12]. The white nodes are good spreaders in terms

of how fast other nodes can be reached from them—this is captured by centrality measures like the Katz prestige. Each shortest path (dashed line) from nodes on the left hand side to the right hand side has to run through the dark-grey node, it has a high betweenness centrality.


page 11 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

New Methods for Deriving RIs: Applications in the Case Studies

We will now consider in-depth examples from three different case studies. In each case we followed the process outlined in section 1.4. To make this transparent, each example is accompanied by a “big data vignette” that provides an overview of (i) the raw data that has been used, (ii) the methodological approach to extract relevant physical or functional networks from that data, and (iii) which kind of RIs (as used in DCLs during assessments with their according IDs) are informed by the data. The overview in this section is by no means complete in the sense that we list here all indicators that are data based, but rather we focus here on those cases where new methodological approaches were required to transform data into indicators, which constituted the core activities of this task. To give a concrete example for case study CHARLIE (healthcare systems), important indicators that are used include the Density of physicians (per 1000 population), or the mortality rate from accidents. These indicators are typically extracted from administrative reporting systems, but there is no need to invoke new methodological approaches in order to compute them. Rather such indicators can be directly measured from the data and the information linked to DCLs in the SCI dashboard, as described in section 4.

2.1 Example I: CHARLIE

Table 1: Big data vignette for CHARLIE.

Steps Description

1. Big and open data used

• Medical claims data o ~8,000,000 patients o ~20,000 healthcare provider o ~300 hospitals with ~2,000 departments o ~200,000,000 contacts with healthcare provider o ~2,000,000 hospital stays

• Transportation data o Accessibility maps based on open street map

(openrouteservice API)1 o Public transportation schedule data from the Wiener Linien

API2 • Population data

o National statistics office and Global Human Settlement framework for spatial information3

2. Methods used Network analysis, geospatially embedded networks, spreading models on networks, isochrones and reachability analysis leveraging GIS information systems, D3 interactive visualization deployment in javascript

1 https://openrouteservice.org/ (accessed Mar 22, 2018). 2 https://www.wienerlinien.at/ogd_realtime/newsList?sender= (accessed Mar 22, 2018). 3 http://ghsl.jrc.ec.europa.eu/ (accessed Mar 22, 2018).

https://openrouteservice.org/

https://www.wienerlinien.at/ogd_realtime/newsList?sender

http://ghsl.jrc.ec.europa.eu/


page 12 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Steps Description

3. Indicator derived

Health coverage indicator (ID-540), Number and rates of disaster-related injuries (ID-521), Number of patients with poisonings (ID-3140), Number of patients with accidents (ID-3150), Population density in potential affected areas (ID-2218), What is the loss in accessibility to health care providers? (ID-3152), What is the percentage of the population that has no access to health care (ID-3153).

The aim of this case study is to assess the resilience of the Austrian healthcare system in scenarios that lead to a surge in the number of patients to be treated within a given time interval for a particular type of healthcare provider, such as hospitals or primary care provider. For instance, a relevant scenario is the flooding of a city and a potentially resulting mass incident of wounded people in an urban area. In such scenarios there is an important dependence of the health infrastructure on the transportation system, as the flooding event may lead to a partial break-down of the transportation infrastructure and therefore put additional stress on certain types of healthcare provider—accessibility becomes an issue [27].

Figure 6: All healthcare providers are shown as circles on a map of Austria (top right), with a zoom into a

region of Vorarlberg (top left). In his zoomed view, the color of the healthcare provider indicates their type (ochre for primary care, turquoise for secondary care or specialists, and blue for hospitals) and the size corresponds to their number of patients treated at a particular day. Note that the positions of the providers have been randomized within a district for reasons of data privacy. It is then possible to follow the movement of individual patients within this map between different healthcare provider (bottom).


page 13 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

For an overview on the data that was used in this case study, see Table 1. The first step of the analysis was to map the healthcare provider in geographical space, see Figure 6. To this end we developed a GIS-based interface that also allowed us to implement interactive features (to be described later). This map shows individual healthcare provider. Within this map, it is then possible to identify the trajectories of individual patients, see Figure 6 (bottom panel). This information is useful for two different purposes. First, it shows that one has to impose an integrated view on the healthcare system, as treatments are typically delivered by a combination of doctors that need to coordinate among themselves. Resilience is a property of the entire health care system, not of individual providers Second, it allows one to build estimates for how patient flows might change should one or several of their providers become inaccessible or inoperational.

Figure 7: Modelling inaccessibility or inoperability of healthcare providers. Left: for each provider the number

of patients is known (here a doctor with 7137 patients is shown) and how any of those patients have also been treated elsewhere (links to other doctors). Right: Should the original doctor be removed, his or her patients are distributed according to those link weights (yellow links). As some providers might now exceed their capacity, a secondary cascade of displacements is triggered (blue lines). The degree of red of the circles shows how close the individual providers are to their capacity.

How this can be done is illustrated in Figure 7 for the example of primary care provider (green circles) in Vorarlberg. To the left there is one provider highlighted with 7,137 patients together with links to other doctors that are proportional to the number of patients that see both of these provider. In the example that is shown, there is one link highlighted that shows that 7% of the patients of the original doctor were also seen by the other one. The idea is that should a specific care provider become inaccessible or inoperational, patient flows are re-routed according to the weights found in this network. For instance, if 50% of the patients that visit hospital A have at some point also visited hospital B, then we formulate a model that assumes that 50% of A’s patients will be transferred to B in case of, say, an evacuation of A. Now, each care provider has a limited capacity. For hospitals this capacity is given by the number of beds and/or doctors in the individual departments (numbers which have been taken from public databases), for other care providers


page 14 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

the capacity can be estimated from historic data (basically as the maximal load of patients that has ever been observed within a given time interval). If the in-flow of patients leads to this capacity being exceeded, patients will be “passed on” again, leading to a secondary cascade of patient displacements. This way a scenario can be defined through the selection of “removed” doctors or care providers and the indicator of interest is how long it takes patients to find a suitable doctor, or how many of them effectively lose access to a specific type of healthcare.

Figure 8: Isochrones are useful to determine the reachability of care providers, such as hospitals. To the left

we show a map of Vienna with markers for places where a hospital is located (here the General Hospital of Vienna is highlighted). To the right we show isochrones for this hospital, shaded areas from which the hospital can be reached in the same time.

As briefly mentioned, a similar modelling approach can be applied to the inpatient and outpatient setting. An important difference for emergency response, however, is that here one is not interested in getting a doctor appointment within a short number of day, but in finding a bed or emergency care provider within minutes. To this end, the “network distance” defined in Figure 7 is less appropriate and rather the travel time to the next hospital becomes of central importance. Such scenarios have been considered with the use of so-called isochrones. The principle is illustrated in Figure 8. To the left we show a map of Austria with markers for each hospital (the General Hospital of Vienna is highlighted). An isochrone for this hospital is defined as follows. For a given time (say, 10 minutes), the isochrones is a line on the map from which it takes 10 minutes to reach the hospital. By changing the considered time span between (e.g., 5 minutes to 2 hours), one obtains a visualization of how fast a specific place can be reached from any other surrounding place. A scenario is then defined by a certain increase in the number of patients that require a specific type of care (say, emergency care in case of a disaster). The General Hospital of Vienna has a staff of maximal 135 doctors and 28 beds available in its emergency department. If the accident occurs in 10 minutes’ distance, this defines the capacity of patients that can be adequately taken care of at the hospital within 10 minutes (and similar for other hospitals and time distances). These measures can now be used to define and measure indicators in the considered scenario, such as the reachability of hospitals, the transport times for the injured population, etc.


page 15 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Figure 9: The models have been deployed using an short interactive throw projector, which allows to define

and play through scenarios by touching the model, as was done here for a demonstration of the technology in front of the Smart City Council of the City of Vienna.

The models described above have been developed so that they can be interacted with on various platforms. They can be deployed using ultra-short-throw interactive throw projectors, which allows users to interface them similar to interactive white board system. That is, initially the map of Austria with all healthcare providers is displayed on a table. By touching the map, a specific region or healthcare provider can be selected to become inoperational or inaccessible (literally by “touching” the provider on the table). The model then shows the resulting patient displacements and gives visual summaries of the result. The idea is that stakeholders can stand together around the table and define and “play” through several scenarios on the fly, see Figure 9. The results of these scenarios are also stored in form of indicators such that they can be uploaded to the SCI dashboard and be used in further assessments.

2.2 Example II: DELTA

Table 2: Big data vignette for DELTA.

Steps Description


• Communication log files of a large scale exercise o ~1,000 participants representing 88 different institutions o ~5,000 time-resolved information flows o 92 simulated events

• Findings replicated and methodology validated on site using a second, similar dataset constructed during an exercise in the following year,

2. Methods used Dynamic network analysis, flow models on networks, regression analysis, multilayer network analysis


Communication between actors (ID-3025); Are there emergency communication channels? (ID-1869) What is the ability to communicate status internally? (ID-3077), What is the ability to internally communicate the status during recovery? (ID-3090).

Communication is a key challenge in the emergency response to disasters, be they man-made or natural. These challenges often belong to one of the following three types: technological, sociological, and organizational [28]. The primary technological challenges include issues arising from the deployment and interoperability of communication systems [29]. Sociological challenges arise from the communication between different groups of people that may use different vocabularies, have different levels of trust towards each other, and different barriers to adopt certain technologies [30]. Finally, organizational


page 16 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

challenges arise when groups that are accustomed to decision-making processes that take place in a vertical hierarchy must suddenly work in a flatter, more horizontal and dynamically changing organization [31]. It has long been observed that under stress, organizations tend to coordinate themselves by feedback instead of planning, effectively leading to a flattening of the hierarchy [32]. Hierarchical organizational structures are known to adapt poorly in highly uncertain environments which often results in information gaps between different organizations [33]. The main challenge tackled here was therefore to quantify to which extent a given organizational structure enables resilient and efficient crisis communication. In particular, we focus on complex emergency situations that require an efficient flow of information between multiple organizations, agencies, and stakeholders. The methodological challenge is posed by the fact that communication networks are typically complex, adaptive, dynamically changing and exhibit non-linear (inter)dependencies [34].

We utilized a unique dataset stemming from the annual public emergency drill by the National University of Public Service in Hungary, see Table 2. During this drill, the participants had to carry out the tasks of their respective organizations under extreme weather conditions that forced them to form an ad-hoc organizational structure to deal with events for which they were either prepared or unprepared. The individual organizations communicated using a peer-to-peer communication system that logged each exchange of information. These log files allow a moment-by-moment measurement of the dynamic communication network of public agencies and organizations in emergency situations, which in turn informed the formulation of indicators for the efficiency and vulnerability of these information flows.

Figure 10: Translating communication network properties into indicators. The clustering coefficient,

betweenness centrality, and Katz prestige can be related to the presence of redundant, vulnerable, and efficient information flows, respectively.

How this formulation of indicators can be achieved is summarized in Figure 10 [35]. Under removal of the highlighted link from A to B, (A), information can no longer be transmitted from A to B if there is no redundant link between them, say via node C (B). The network property of having such redundant links can be measured by the clustering coefficient of a network. The removal of the node highlighted in (C) does not change the ability of all other nodes to exchange information, whereas in (D) the nodes on the left become disconnected from those on the right, which indicates a substantial vulnerability. This property can be measured by betweenness centrality. As opposed to the case shown in (E), where information can only flow in one direction, the existence of feedback loops, (F), allows information to be efficiently transmitted between each pair of nodes in the network. This property can be quantified using eigenvalue-based indicators, such as Katz prestige.

The task of the exercise carried out on 18-19th of April, 2016, was to organize and lead the activities of public service organizations among extraordinary weather conditions. The simulated time lapse was from 2016.12.18 to 2016.12.24. During this simulated time, the weather situation changed dynamically with the


page 17 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

arrival of a cyclonic storm with freezing rain and cold (Dec 18), to a strong breeze and temperatures down to -34°C (Dec 19), that continued until Dec 23. A month before the exercise, the students learnt about the theme and the objectives of the exercise. They learnt about the organization in which they would work as a member and the specific role (position) in which they would need to carry out their tasks. They were also given the required and recommended reading material before the exercise, having a one-month preparation period during which they were assisted by teachers. The exercise consisted of two types of events, prepared and unprepared ones. Prepared events were those for which the participants where explicitly briefed beforehand. Such events were expected to lead to “vertical” information flows between different hierarchical levels (e.g., information being passed between local and national units). In addition, there were unprepared events that represented emergency situations due to the weather crisis (for instance, a blocked road). Such events where expected to lead to “horizontal” information flows, as local police units had to coordinate with firefighters, ambulances, road authorities, and even the military to request airlift for rescue and caterpillars to reach the site off-road.

Figure 11: Communication networks for all messages sent on day 1 (left) and for day 2 (right). Each node

corresponds to one organization, directed links between two nodes indicate information flows. Node colors correspond to hierarchical levels, i.e., blue for the high (national or regional) and green for low (local) levels. Node size is proportional to the number of sent messages. Both networks consisted of multiple communities of nodes. However, on day 2 there was a much higher number of information flows between nodes from different communities all across the network as compared to day 1

Prepared events took place on the first day of the exercise (real time, not simulated time), whereas unprepared ones occurred during the second day. The communication networks for these two respective days are shown in Figure 11. Nodes in the networks correspond to individual organizations, a directed link between two nodes indicates a message sent from the source to the target node. Node color indicates the hierarchical level of the given node (blue for the national/regional and green for the local level). Node size is proportional to the number of sent messages of the corresponding organization. On both days there were well discernible groups of densely interconnected nodes, so-called communities [36]. There were two communities of nodes on the higher hierarchical level on each day, and a central community on the lower hierarchical level that corresponded to local authorities. There were several peripheral groups of nodes on the lower level, too. By comparing the networks for day 1 and 2, it can be seen that the number of links that connected different groups of nodes was much higher on day 2 than on day 1. That is, the participants had an increased need to coordinate themselves all across the network on day 2 when they were unprepared.


page 18 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

For each of the 92 simulated events (prepared and unprepared ones) we computed the indicators shown in Figure 10 for the networks of communication flows that were observed directly after the events. This allows to quantify the communication behavior of each participating organization individually, as well as for the entire network of organizations. As a proof-of-principle result, we show these results visualized through a 3d scatter plot in Figure 12 (indicators on network-level). Here each point corresponds to a single observation of the communication network. Red circles show networks of prepared participants; blue squares correspond to unprepared ones. There is a clear tendency for the red data points to be shifted towards higher levels of efficiency, lower vulnerability, but also lower redundancy.

Figure 12: 3d scatter plot of the network indicators that represent efficiency, redundancy, and vulnerability.

Each point corresponds to the communication network at a given time with participants being either prepared (red circles) or unprepared (blue squares). There is a clear tendency that prepared participants showed higher efficiency, lower vulnerability and less redundancy in their communication.

We now turn our attention from indicators for the entire communication network to the redundancy, vulnerability, and efficiency properties of individual organizations. In particular, we were interested in whether there were systematic differences between the two hierarchical levels – high (national, regional) and low (local). We computed the network indicators for day 1 and day 2, respectively and compared them using statistical testing procedures. We found for the unprepared case an increased level of individual redundancy on the lower hierarchical level, but not on the higher level. There was a strongly significant decrease in efficiency on the higher level as the participants became unprepared, but no significant change on the lower level. In terms of vulnerability we found no significant changes on both levels, but an overall weak trend towards increased vulnerability. In other words, the analysis has confirmed the intuitive, “gut” feeling, that when the personnel were prepared for a given event (e.g., critical weather condition), vertical communication within the organization is efficiently at work. However, when sudden events occurred, the affected units concentrated on horizontal communication, thereby creating bottlenecks and loss of efficiency on the systemic level of the entire network. The indicators presented here enable us to quantify this qualitative statement and to link it to real organizations and the people acting within them.

To further validate the methodology of deriving and applying indicators for communication during emergency response, we repeated the analysis during the 2017 emergency drill—but this time on site using live data. We manned a desk during the drill where analysts had access to the communication data and could extract the indicators as the drill unfolded. Examples for the output of this analysis are shown in Figure 13, which shows a snapshot of the communication network during a specific time interval, colors this time representing the vulnerability and efficiency indicators, respectively. This information was used during de-briefing in order to identify participants that did not fulfill their expected roles in establishing information flows. With the lessons learned from this live exercise, we are currently exploring the application of similar


page 19 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

methods during the DELTA drill, but this time also including a layer of geospatial information from the participants’ RFID tags.

Figure 13: Communication networks extracted on site with live data during the 2017 emergency drill. The

node colors indicate organizations with high vulnerability (left) and efficiency (right), respectively. This information was used during the de-briefing of the exercise.

2.3 Example III: ECHO

Table 3: Big data vignette for ECHO.

Steps Description


• Sensor data from a petrochemical plant o 954 sensors o 46,081 time intervals o ~44 million data points

2. Methods used Dynamic network analysis, correlation network analysis, time series analysis, trend detection, regression models, principal component analysis


Big data analyzed (ID-2990); Fresh steam flow to the rec. charge (ID-2992); Turbine exhaust pressure (ID-2993), Lube oil temperature behind cooler (ID-2994); Return LCGO reflux (ID-2995); Early Warning Signal Level derived from Big Data (ID-3059).

The aim of the case study ECHO is to assess functionality and resilience of an interconnected industrial production system with specific application to a petrochemical plant—the NIS Pančevo Oil Refinery. The main reason for this exercise was to show how the SmartResilience approach can further enhance the risk management operations already in place at NIS. To this end NIS provided access to time series data from about 1,000 sensors in the petrochemical plant, see Table 3. These measurements covered a timespan of one month that included an adverse event, namely an output quality deterioration that eventually led to a cessation of operations at the plant. The challenge posed was therefore to derive RIs that allow one to identify signals with greatest importance using heterogeneous types of dynamic information, and in particular how precursors for such adverse events could be detected as early as possible.


page 20 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Figure 14: Overview of workflow of the big-data indicators in ECHO. Based on the sensor data NIS Pančevo

Oil Refinery, a statistical time series analysis was performed to detect outlier measurements. Whether such outliers signal that other components might be affected as well is determined in a correlation network analysis that informs early warning indicators. These indicators can be used in DCLs for short-term monitoring.

The workflow of deriving and using the RIs is summarized in Figure 14. In brief, the data provided by NIS was (i) subjected to a time series analysis in which sensors and time points of “unusual” or outlier measurements were identified, (ii) followed by a network analysis to identify “critical” measurement fluctuations which can be formulated in terms of (iii) and early warning indicator that can be readily used in (iv) DCls for monitoring purposes. An alarm value can then be set for these indicators in order to inform operators in case an early warning signal is detected in the data. Steps (i)-(iii) constitute the methodological core of the analysis, whereas the other steps are implemented in an integrated tool—the SCI dashboard—and will be further described in section 4.

Figure 15: Deriving RIs based on sensor time series. Left: anomalies in the data of individual sensors can be

detected by measuring by how many standard deviations the current measurement exceeds the


page 21 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

values normally observed. Outliers are shown as red spikes. Right: Whether such an outlier is a precursor that further deviations should be expected or not can be determined from the functional network of the sensors. Here, each node represents a sensor and two sensors are linked if the measurements at these sensors typically correlate. Early warnings are those outliers that occur at nodes with a high centrality in the functional network (shown as the degree of red of the node colors). Note that the functional network has well discernible modules (two of them being highlighted by green circles) that closely mimic the actual physical layout of the plant.

Let us briefly review the core methods in Steps (i)-(iii). For the time series analysis, step (i), a so-called Z-transform is applied to each time series, i.e., the data is rescaled such that each time series has the same average value and—on average—shows fluctuations with the same strength (unit standard deviation). At a certain point in time and for a specific sensor, anomalies can then be detected based on by how much standard deviations the current measurement exceeds the values usually observed, see Figure 15. At this point, however, it is not clear whether such an outlier might signal that further deviations from the normal mode of operation are to be expected, or whether the outlier is likely to be a singular event without any further consequences. Therefore, the next step (ii) was to consider the functional network of sensors by means of a correlation network analysis. In this network each node corresponds to a sensor and two sensors are linked if fluctuations recorded by one of them typically coincide with fluctuations in measurements of other sensors. This network is shown in Figure 15. We see that there are tightly interconnected modules of components that also share connections among themselves. We have verified with the infrastructure operator that this structure of the network indeed follows the layout of the plant (which is information that was not included in the analysis beforehand, but that is rather extracted in the form of a functional network architecture). The rationale is then that an outlier in a sensor that is central in the functional network is more reason for concern than a fluctuation in a sensor in the periphery of the network (where only very little other components of plant will “feel” the fluctuation). As a measure of centrality of a sensor in the network, we again use Katz prestige. By weighting the fluctuations observed in a specific sensor by its centrality in the functional network of the plant, an early warning indicators can be derived, step (iii).

As can be seen in Figure 15, there are many outliers of quite large magnitude occurring over time. However, there is a clear accumulation of outliers towards the end of the observation window. This is a clear signal that the anomaly detection senses the impending output quality deterioration. More concrete, we show the time series for the early warning indicator with the strongest observed signal in Figure 16. For the first couple of weeks there was no signal whatsoever, followed by a strong signal just before the end of the data, where the deterioration occurred. The indicator therefore behaves in a way that one would expect from a meaningful early warning signal.

Figure 16: Early warning indicator for the event observed in the data. The indicator for the sensor shown

here shows no signals until briefly before the end of the observation window where the output quality deteriorated, where a strong signal is picked up. This event was indeed the strongest observed event in the data, as shown by the ranks of detected early warnings.


page 22 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

2.4 Example IV: GOLF

Table 4: Big data vignette for GOLF.

Steps Description

1. Big and organisational data used

• Callout data Cork Fire Brigade 2013 - 2017 o more than 13,000 datapoints

• Transportation data o matrices data describing traffic flow on street level based on

demand models

2. Methods used time series analysis

3. Indicator derived Resource allocation of first responders

Traffic statistics for each street in the city

The GOLF case study aims to assess available datasets for their application in a resilience context of a flooding disaster case study. We reviewed data provided by Cork City Council and the National Transport Authority.

Being able to answer questions such as how many first responder resources are typically occupied responding to an incident at certain times of a day can help optimise planning for additional staff to be available to respond to predicted flooding of the city on top of expected day by day, hour by hour incident volume. Further being able to understand traffic demand on a street level allows planning for adequate response strategies such as media alerts, resource planning to address an expected high number of civilians in certain parts of the city, and a high street traffic volume impacting emergency response that will need to use specific streets.

We assessed the callout data for insight on the use of first responder resources. We created matrices describing the callouts

- per month in each year - per hour per day of the week

Using the matrices, we carried out a trend analysis which gave insight into the use of first responder resources which can help with resource planning meeting the incident volume per time frame.

The National Transport Authority (NTA) has developed a system of regional models to provide a detailed multi-modal, state-of-the-art transport assessment toolkit for each of the five main city-regions in the Republic of Ireland. The system contains a detailed representation of the transport networks in the urban centres, the fine zoning system in these areas enables the Demand Model component to accurately represent travel patterns in these areas. The Cork City matrices data allows to get the following insight per street level during different times of the day:

- free flow speed - congested speed - number of lanes - bus lanes

2.5 Indicators for other case studies Besides the in-depth examples given above, big or open data related indicators are also used in the other case studies of the projects. In many of these cases, indicators could be formulated for the relevant datasets (see [1]) in a quite straight-forward way, such as counts of certain events. In other cases, the indicators are based on data analysis methods that are well known and described in the respective fields of applications. To give a brief overview of these activities, in Table 5 we show a selection of indicators that are currently used


page 23 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

or considered for DCLs in various case studies. More information on how the indicators are defined and how they are measures can be found in the SCI dashboard by looking them up using their IDs. As various tasks in WP5 are still underway, the list in Table 5 makes no attempt at being complete.

Table 5: Selection of big and open data related indicators across different case studies.

Indicator ID Indicator Name Case study

ID-2848 Forensic data recovery expertise ALPHA

ID-2843 Percentage of companies buying cyber insurance ALPHA

ID-3276 Average shutdown time in minutes BRAVO

ID-1121 Is the power grid robust by means of the n-1 stability criterion? BRAVO

ID-2402 Percentage of households supplied with electricity BRAVO

ID-1045 Average capacity of the electric power backup t support the continued operations? BRAVO

ID-575 How many cyber security breaches have occurred in the past year? BRAVO

ID-578 How many unplanned shutdowns have occurred in the past year? BRAVO

ID-607 Is security data tracked automatically? BRAVO

ID-606 Can security data be analyzed in real time BRAVO

ID-631 Is the overall system structured in a mesh BRAVO

ID-521 Number and rates of disaster-related injuries CHARLIE

ID-3140 Number of patients with poisonings CHARLIE

ID-3150 Number of patients with accidents CHARLIE

ID-2218 Population density in potential affected areas CHARLIE

ID-2182 RFID based indicator DELTA

ID-2183 CCTV cameras data DELTA

ID-2184 HD cameras with image analysis software (face recognition, person counting) DELTA

ID-2178 Cargo payload from the bills of material database DELTA

ID-2186 Number of cargo flights takeoff rate (flights per hour) DELTA

ID-2187 Number of flights per hour DELTA

ID-2188 Number of registered incidents per hour DELTA

ID-2189 Updated number of incidents per 1 million passengers DELTA

ID-1057 Recording of the recovery time ECHO

ID-2072 The existence of a register of accidents/incidents ECHO

ID-1235 Domestic oil production (thousand tons/day) ECHO

ID-1237 Refining of domestic crude oil (thousand tons/day) ECHO

ID-1238 Refining of crude semi-finished products (thousand tons/day) ECHO

ID-1271 Economic inoperability between NIS Serbia and other countries (€/day) ECHO

ID-1273 Exports (thousand tons/day) ECHO

ID-2191 Total income from all activities (€) ECHO

ID-1894 What is the level of energy consumption for water treatment? FOXTROT

ID-1831 How often do representatives (drinking water) assess data from healthcare? FOXTROT


page 24 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Indicator ID Indicator Name Case study

ID-1829 What is the current drinking water consumption per capita? FOXTROT

ID-1813 Raw water chemical quality (Issue with indicators IDs 1922-1939) FOXTROT

ID-3158 Percentage of population that is symptom free FOXTROT

ID-3155 Percentage of functionality in SCADA and instrumentation FOXTROT

ID-3236 Are processed and stored data defined? GOLF

ID-608 Can security data be analyzed in real time? GOLF

ID-3252 Are data and meta-data available? GOLF

ID-2200 Number of tidal and river gauges in places GOLF

ID-2225 Number of OPW Flood Maps available GOLF

ID-2263 Is there a system in place for data analysis after crisis? GOLF

ID-2889 Number of automatic river level gauges GOLF

ID-2901 Availability of weather forecast data GOLF

ID-2885 Lack of critical data models (issue with indicator IDs 2863, 2902, 2903) GOLF

ID-2977 Flood gauges and rainfall monitoring data GOLF

ID-2960 Number of followers on alerting systems GOLF

ID-2962 Measure social media as a percentage of total question / inquiry volume GOLF

ID-2344 Production for district heating (percentage of the full normal capacity) HOTEL

ID-3167 Impact of lost revenue HOTEL

ID-2409 Revenues (% of production at full capacity) HOTEL

ID-2345 Interconnectedness of district heating system (issue) HOTEL


page 25 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Linking the indicators: INDIA and interdependences

Table 6: Big data vignette for interdependences.

Steps Description


• EUROSTAT data tables, including GDP and main components, symmetric input-output tables, fixed assets by industry, balance sheets of non-financial assets, non-financial transactions, business demography by legal form, capacity utilization in manufacturing, government revenue, deficit/surplus, debt and expenditure, population by economic activity, money market interests

• Census data from national statistical offices • Demographic surveys • National account data (governmental and sectoral) • Business surveys • European Labour Force Survey • SABINA database of financial statements • World Input-Output Database

2. Methods used Agent-based modeling, inoperability input-output models, environmental damage models, network analysis


In principle all indicators qualify as indicators for interdependences, given that they are either (i) evaluated by considering “infrastructures of infrastructures” or (ii) used within an issue that deals with dependence on one infrastructure of an adverse event in another one.

Big and open data based indicators were also an essential part of the interdependence analysis carried out in task 2.3, and in particular in the economic and agent-based models developed there [37]. Here, we do not discuss these efforts in detail, which was done in the corresponding deliverable, but simply report the corresponding “big data vignette”, see Table 6.


page 26 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Figure 17: Overview of the indicator-based approach to interdependences. The case-study infrastructures are

shown as large green circles, indicators as small blue circles. Links connect infrastructures and indicators if the indicator has been used in the assessment of an infrastructure.

A different data-intensive approach to interdependence that was described in the deliverable for task 2.3 is the linking of issues and indicators through their concurrent use in the same assessments (“indicator-based approach”). To briefly recapitulate this approach, see Figure 17. This approach will play an important role in linking the results (and indicators) from individual case studies, in particular within the integrated case study INDIA. The idea is to perform an integrated assessment of “infrastructures of infrastructures” within INDIA, including cascading and ripple effects between them, by linking the infrastructures through common issues and indicators. This way, a combined DCL can be formulated. To ensure the success of this approach, it is necessary that the case study leaders provide their input to their respective tasks in WP5 in a structured and consistent form. In particular, the following requirements must be met.

• All assessments in the individual case studies are conducted through the SCI dashboard. • The case study leaders preferentially use approved indicators and issues from recommended or

core DCLs. That is, it should be avoided that each case study defines its “silo” of indicators and issues, resulting in the same things being measured under different names within each individual case study.

• It follows that new issues and indicators should only be introduced to the database in case an exhaustive search has been performed that shows that no such issues or indicators have already been used.

Adherence to the above requirements should ensure that meaningful overlap between individual indicators and issues is generated through their shared use. Based on this shared use, standard recommendation system approaches can be used to make automated predictions on useful indicators for individual users. In this context, a promising standard technique is collaborative filtering [38], i.e., a method of making automated predictions for a particular user based on the collected information from a large number of other, similar users. The rationale behind this approach is that if a user is evaluating a particular scenario and/or particular issues, then the indicators that have been used in other assessments of the same scenario are relevant. In the context of indicator recommendations this means the following. The use of a particular


page 27 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

indicator by a user in a DCL can be interpreted as a preference to use this indicator when assessing a particular threat, phase, dimension, issue, and infrastructure. The more often it is observed that the same indicator is used by many different users when assessing the same threat, phase, dimension, issue, or infrastructure, the more confidence we can have in recommending the considered indicator to other users in similar assessments. That is, once a user has specified, for instance, an infrastructure, a threat, and an issue, we recommend those indicators that other users often used in similar assessments. These indicators can be identified from networks similar to the one shown in Figure 17. For each use of an indicator for a given infrastructure, we increase the weight of the corresponding link (and similar for other items like issues, threats, etc.). Recommendations can then be built by identifying the indicator-nodes that have links with the strongest weights to the items that have been selected by the user so far.


page 28 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Using big data indicators in resilience assessments—the SCI dashboard

In all the examples we have discussed so far in this report we encounter the following challenge. Within the project, more or less sophisticated mathematical and algorithmic network-based forecasting and simulation models have been developed to derive RIs. These models typically require a quantitative analyst in order to be effectively implemented. This need to implement such models would render the turn-around in simulating scenarios rather slow or even impossible, as not every end user might have quantitative analysts at his or her disposal (particularly when considering SMEs). It is therefore no option to manually re-code the methods developed here in order to deploy them at a specific end user. The solution to this challenge that we pursue in the SmartResilience project is therefore to deploy the models to a web-based interface such that other users can run and interact with them in various assessments.

The corresponding workflow for the end user is shown in Figure 18. The first step for the end user is to provide input in the form of specifying a DCL. If desired, the user can choose to upload data as described in this report to the SCI dashboard. Therefore, an Excel template is provided within the tool that contains instructions on the expected data format. Once the data has been uploaded to the database, analyses can be carried out through a web-based interface that provides access to libraries (DLL files) that execute the computations. These DLL files can, for instance, consist of deployed MatLab solutions that expose certain functions that compute the indicators based on the input data, i.e., functions that extract the relevant networks and compute performance measures such as Katz prestige. For the user, these computations take place entirely in the background. The output that he or she receives are assessments (of functionality or resilience that also include conventional indicators).


page 29 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Figure 18: SmartResilience use case workflow. The end user provides specifications of a DCL as input and

uploads his or her data. The SCI dashboard offers access to web-based deployments of the big data analytics services described in this report, which occur in the background. The output for the end user is an assessment that includes data-intensive and conventional indicators.

To be more concrete, the steps to use data analytics background service are as follows, see also Figure 19. The first step for the user is to login to the SmartResilience member-only area and go the SCI dashboard. Then the scenario needs to be designed and the DCL needs to be created. Therefore, the user selects the CI, issues, indicators, and assigns values where to the indicators whenever this input is available. To utilize the data analytics background service, the next step is to select those indicators in which the value should be derived from the network analysis approaches described in this report, which are all included as existing indicators. This leads to the next step, where the corresponding templates can be downloaded. Once the user has inserted his or her data into this template, it can be uploaded again and the analysis is run with the click of a button. This takes the user to the output—an assessment based on big data and/or conventional indicator.


page 30 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Figure 19: Illustration of workflow in the SCI dashboard to utilize the big data indicators. After the DCL has

been generated, existing big data indicators can be selected. A template is then offered that allows the user to upload his or her data and run the analysis with the click of a button. In the case shown here, the result is an assessment of the resilience level in the ECHO case study.


page 31 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Conclusions

In this report we have described work toward deriving RI from big and open data. Our starting point was that modern critical infrastructures—in particular smart infrastructures—routinely produce abundances of data that can be leveraged in its resilience and functionality assessment. The central challenge in doing so is that this data is typically dynamic, heterogeneous, interconnected, and/or multirelational, which means that simple approaches of data analysis that do not take these characteristics properly into account are often of limited use. Instead, networks naturally arise in the formal conceptualization of such systems, be it that the infrastructures can be represented as actual, physical networks (transmission or transportation networks) or that functional networks can be extracted from the data. We therefore developed a methodological approach in which the identified datasets are (i) mapped onto networks, (ii) characterized in terms of network properties that (iii) can be related to resilience indicators.

We provided several working examples within case studies in the project where these steps have successfully been implemented. This includes the case study CHARLIE, that analyses the resilience of a national healthcare system with respect to adverse events that lead to a reduction of (accessible or operational) healthcare providers in a specific region. By using medical claims data (which is by nature highly dynamic, interconnected, heterogeneous, and multirelational), combined with other open data sources, we could derive networks of patient flows between individual healthcare provider that can be used to model how well individual provider (hospitals, doctors) or entire regions can absorb sudden surges in the number of patients to be treated. The models have been deployed using state-of-the-art interactive whiteboard systems, that allow to collaboratively use such models in generating and playing through scenarios.

Another example was provided in the case study DELTA, where we analyzed the entirety of communication events that have been recorded during a large public emergency drill that required the coordination of thousand participants across several organizations (dynamic, heterogeneous and interconnected data). We showed how indicators can be derived that are informative on the efficiency, vulnerability, and redundancy of individual organizations, but also for how well these organizations coordinate among themselves. This approach was later tested on-site during the emergency drill in the following year, which showed that the methods also work using live data such that the results of the analysis were available during debriefing.

Case study ECHO demonstrated how early warning indicators for adverse events can be derived from finely time-resolved sensor data from a petrochemical plant (data that is again dynamic, heterogeneous and interconnected). Therefore, we extracted the functional network of sensors that shows how likely a perturbation in one component of the plant is to spread to other components. By combining this information with anomaly detection in the sensor time series, we showed how indicators can be derived that signal output quality deterioration.

Big data analysis also played a key part when addressing interdependences of infrastructures, see also task 2.3. We discussed how the data generated in the project itself—namely the combined information of all resilience assessments that are carried out in the project—can be used to address the assessment of “infrastructures of infrastructures”. The idea is to build recommendation systems based on collaborative filtering to automatically prepare DCLs for the joint assessment of multiple infrastructures.

Finally, we showed that the approach to derive RIs from big and open data is fully integrated within the SmartResilience tools—the SCI dashboard. This allows end users to leverage the network-based analysis without the need to implement or deploy the analysis models themselves. Instead, after a DCL has been generated, big data indicators can be selected that provide a template for the user to upload his or her own data. A data analytics background service then carries out the analysis and the user receives the resilience or functionality assessment with a couple of mouse clicks.


page 32 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

References

[1] SmartResilience (2016), Deliverable D5.1 Report on the results of the interactive workshop. http://smartresilience2.eu-vri.eu/filehandler.ashx?path=SmartResilience/Submitted%20deliverables/D5.1-InteractiveWS_SUBMITTED.pdf

[2] Boccaletti, Stefano, et al. "The structure and dynamics of multilayer networks." Physics Reports 544.1 (2014): 1-122.

[3] Kivelä, Mikko, et al. "Multilayer networks." Journal of complex networks 2.3 (2014): 203-271. [4] Little, Richard G. "Controlling cascading failure: Understanding the vulnerabilities of

interconnected infrastructures." Journal of Urban Technology 9.1 (2002): 109-123. [5] Wang, Wen-Xu, and Guanrong Chen. "Universal robustness characteristic of weighted networks

against cascading failure." Physical Review E 77.2 (2008): 026101. [6] Goldenberg, Anna, et al. "A survey of statistical network models." Foundations and Trends in

Machine Learning 2.2 (2010): 129-233. [7] Menard, Scott. Applied logistic regression analysis. Vol. 106. SAGE publications, 2018. [8] Newman, Mark. Networks: an introduction. Oxford university press, 2010. [9] Corsi, S., and C. Sabelli. "General blackout in italy sunday september 28, 2003, h. 03: 28: 00."

Power Engineering Society General Meeting, 2004. IEEE. IEEE, 2004. [10] Buldyrev, Sergey V., et al. "Catastrophic cascade of failures in interdependent networks." Nature

464.7291 (2010): 1025. [11] Callaway, Duncan S., et al. "Network robustness and fragility: Percolation on random graphs."

Physical review letters 85.25 (2000): 5468. [12] Thurner S, Klimek P, and Hanel R. Introduction to the Theory of Complex Systems. Oxford

university press, 2018. [13] Haldane, Andrew G., and Robert M. May. "Systemic risk in banking ecosystems." Nature

469.7330 (2011): 351. [14] Acemoglu, Daron, Asuman Ozdaglar, and Alireza Tahbaz-Salehi. "Systemic risk and stability in

financial networks." American Economic Review 105.2 (2015): 564-608. [15] Battiston, Stefano, et al. "Debtrank: Too central to fail? financial networks, the fed and systemic

risk." Scientific reports 2 (2012): 541. [16] Poledna, Sebastian, et al. "The multi-layer network nature of systemic risk and its implications for

the costs of financial crises." Journal of Financial Stability 20 (2015): 70-81. [17] Langfelder, Peter, and Steve Horvath. "WGCNA: an R package for weighted correlation network

analysis." BMC bioinformatics 9.1 (2008): 559. [18] Pearson, Karl. "LIII. On lines and planes of closest fit to systems of points in space." The London,

Edinburgh, and Dublin Philosophical Magazine and Journal of Science 2.11 (1901): 559-572. [19] Noble, William S. "How does multiple testing correction work?." Nature biotechnology 27.12

(2009): 1135. [20] Serrano, M. Ángeles, Marián Boguná, and Alessandro Vespignani. "Extracting the multiscale

backbone of complex weighted networks." Proceedings of the national academy of sciences 106.16 (2009): 6483-6488.

[21] Pastor-Satorras, R., Vespignani, A. “Epidemic spreading in scale-free networks.” Physical review letters 86.14 (2001): 3200.

http://smartresilience2.eu-vri.eu/filehandler.ashx?path=SmartResilience/Submitted%20deliverables/D2.2-Challenges_v12bc24022017.pdf




page 33 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

[22] Elmasri, Ramez, and Shamkant Navathe. Fundamentals of database systems. Addison-Wesley Publishing Company, 2010.

[23] Brandes, Ulrik. "A faster algorithm for betweenness centrality." Journal of mathematical sociology 25.2 (2001): 163-177.

[24] Bonacich, Phillip. "Power and centrality: A family of measures." American journal of sociology 92.5 (1987): 1170-1182.

[25] Katz, Leo. "A new status index derived from sociometric analysis." Psychometrika 18.1 (1953): 39-43.

[26] Page, Lawrence, et al. The PageRank citation ranking: Bringing order to the web. Stanford InfoLab, 1999.

[27] Lo Sardo R, et al. “Systemic risk in healthcare networks”. In preparation (2018). [28] Manoj, Balakrishan S., and Alexandra Hubenko Baker. "Communication challenges in emergency

response." Communications of the ACM 50.3 (2007): 51-53. [29] Braunstein, Brian, et al. "Feasibility of using distributed wireless mesh networks for medical

emergency response." AMIA annual symposium proceedings. Vol. 2006. American Medical Informatics Association, 2006.

[30] Tierney, K., and J. Sutton. "Cost and culture: Barriers to the adoption of technology in emergency management." RESCUE Research Highlights (2005).

[31] Botterell, Art, and Martin Griss. "Toward the next generation of emergency operations systems." (2011).

[32] Dynes, Russell R., and Enrico Louis Quarantelli. "Organization communications and decision making in crises." (1976).

[33] Palttala, Pauliina, et al. "Communication gaps in disaster management: Perceptions by experts from governmental and non-governmental organizations." Journal of Contingencies and Crisis Management 20.1 (2012): 2-12.

[34] Wasserman, Stanley, and Katherine Faust. Social network analysis: Methods and applications. Vol. 8. Cambridge university press, 1994.

[35] Klimek P, et al. “Resilience Indicators for Information Flows in Organizational Communication Networks in Emergency Response”. In review (2018).

[36] Fortunato, Santo. "Community detection in graphs." Physics reports 486.3-5 (2010): 75-174. [37] SmartResilience (2018), Deliverable 2.3 Report on interdependencies and cascading effects of

smart city infrastructures. [38] Linden, Greg, Brent Smith, and Jeremy York. "Amazon. com recommendations: Item-to-item

collaborative filtering." IEEE Internet computing 7.1 (2003): 76-80.


page 34 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

A N N E X E S


page 35 Smar

tRes

ilien

ce -

D4.

2 Re

silie

nce

indi

cato

rs fo

r SCI

s ba

sed

on b

ig a

nd o

pen

data

Annex 1 Review process

The Content of this Annex has been submitted as part of the periodic review report to the PO/EU/ Reviewers.

smartresilience - d4.2 resilience indicators for scis ... · assessing resilience of scis (#2) by...

Documents