Detection of Damage and Failure Events of Critical Public

Infrastructure using Social Sensor Big Data

Iris Tien

, Aibek Musaev

, David Benas

, Ameya Ghadi

, Seymour Goodman

and Calton Pu

School of Civil and Environmental Engineering, Georgia Institute of Technology,

790 Atlantic Drive, 30332-0355, Atlanta, GA, U.S.A.

School of Computer Science, Georgia Institute of Technology, Atlanta, GA, U.S.A.

Sam Nunn School of International Affairs, Georgia Institute of Technology, Atlanta, GA, U.S.A.

Keywords: Social Sensors, Big Data, Data Processing, Critical Infrastructure, Event Detection.

Abstract: Public infrastructure systems provide many of the services that are critical to the health, functioning, and

security of society. Many of these infrastructures, however, lack continuous physical sensor monitoring to

be able to detect failure events or damage that has occurred to these systems. We propose the use of social

sensor big data to detect these events. We focus on two main infrastructure systems, transportation and

energy, and use data from Twitter streams to detect damage to bridges, highways, gas lines, and power

infrastructure. Through a three-step filtering approach and assignment to geographical cells, we are able to

filter out noise in this data to produce relevant geolocated tweets identifying failure events. Applying the

strategy to real-world data, we demonstrate the ability of our approach to utilize social sensor big data to

detect damage and failure events in these critical public infrastructures.

1 INTRODUCTION

Public infrastructure systems provide many of the

services that are critical to the continued health,

functioning, and security of society. This includes

energy systems that power nearly all devices,

controls, and equipment, as well as transportation

systems that enable the movement of people and

goods across both short and long distances. Failure

of or damage that has occurred to these

infrastructures, whether from deterioration and

aging, or from severe loads due to hazards such as

natural disasters, poses significant risks to

populations around the world. Detecting these

damage or failure events is critical both to minimize

the negative impacts of these events, e.g., by

rerouting vehicles away from failed bridges, and to

accelerate our ability to recover from these events,

e.g., by locating the extent of power outages for

deployment of repair crews.

Many of these infrastructures, however, lack

continuous physical sensor monitoring to be able to

detect these damage or failure events. Bridges, for

example, are generally subject to only yearly

inspections, and very few are instrumented with

physical sensors that would be able to detect damage

that may occur at any time. In addition,

infrastructures that contain monitoring capabilities,

such as energy systems, may have extensive

networks of physical sensors at a centralized level,

but less so at the distribution level. Thus, while

power plants are closely monitored, maps of outages

rely on individual reports.

In this paper, we propose the use of social

sensors to detect damage and failure events of

critical public infrastructure. Recently, there has

been an exploration of the use of data from social

sensors to detect events for which physical sensors

are lacking. This includes the use of Twitter data

streams to detect natural disasters (Sakaki et al.,

2010) or the use of texts to manage emergency

response (Caragea et al., 2011). In this paper, we use

the LITMUS framework – a framework designed to

detect landslides using a multi-service composition

approach (Musaev et al., 2014a, 2014b) – to detect

public infrastructure failure events. We focus on two

main systems: transportation (bridges and highways)

and energy (gas lines and power).

The rest of the paper is organized as follows.

Section 2 provides an overview of the approach used

Tien, I., Musaev, A., Benas, D., Ghadi, A., Goodman, S. and Pu, C.

Detection of Damage and Failure Events of Critical Public Infrastructure using Social Sensor Big Data.

DOI: 10.5220/0005932104350440

In Proceedings of the International Conference on Internet of Things and Big Data (IoTBD 2016), pages 435-440

ISBN: 978-989-758-183-0

435

to detect infrastructure failure events using social

sensor data. Section 3 provides the results of the

approach as applied to four infrastructures: bridges,

highways, gas lines, and power. In Section 4, we

provide an evaluation of the proposed approach,

including filtering results for the social sensor data

and visualizations of the detected events. We

summarize related work in Section 5 and conclude

the paper in Section 6.

2 APPROACH

An overview of the approach is shown in Figure 1.

The sensor data source is Twitter. For the results

presented in this paper, these are tweets pulled over

the period of one month. We use October 2015 as

our evaluation period. It is noted that data from any

other time period can be used within this framework.

To detect infrastructure damage or failure events,

all Twitter data is run through a series of filters to

obtain a subset of relevant data. This filtering is

done in three phases. First, we filter by search terms,

which we have developed for various events of

interest, e.g., “bridge collapse” to detect damage to

bridge infrastructure. Second, as social sensor data is

often noisy, with items containing the search terms

but unrelated to the event of interest, data is filtered

using stop words. Using a simple exclusion rule

based on the presence of stop words, this filters out

the irrelevant data. An example for detecting bridge

collapses is the stop word “friendship” that refers to

the collapse of a bridge or connection between two

people.

Third, data is filtered based on geolocation.

Although most social networks enable users to

geotag their locations, e.g., when they send a tweet,

studies have shown that less than 0.42% of tweets

use this functionality (Cheng et al., 2010). In

addition, users may purposely input incorrect

location information in their Twitter profiles (Hecht

et al., 2011). As geolocating tweets is an important

component in being able to identify specific

infrastructure damage events, including their

location, the data must be additionally filtered. In

this study, the Stanford coreNLP toolkit (Manning et

al., 2014) is used along with geocoding (Google,

2016) to geolocate the tweet. This assigns each

filtered tweet to a latitude and longitude and

corresponding 2.5-minute by 2.5-minute cell as

proposed in Musaev et al., 2014, based on a grid

mapped to the surface of the Earth.

Once all relevant tweets are mapped to their

respective cells, all tweets in a single cell are

assessed to identify the infrastructure damage and

failure events. In this paper, we focus on the results

for tweets relating to damage detection in four

infrastructures: bridge, highway, gas line, and power

infrastructure.

Figure 1: Overview of data, filtering, and event detection

approach.

3 RESULTS

Each of the four infrastructures studied present

different challenges, with particular characteristics

for filtering that we discuss in this section. In

addition, we present the specific search terms and

stop words that we have found for use in identifying

events of interest for each infrastructure. All Twitter

data is filtered using these words to obtain the subset

of relevant data, which is then geolocated to identify

the damage or failure events.

3.1 Bridges

Bridge-related damage events tend to be major

events. This includes closures of bridges that are part

of major transportation arteries, or high-visibility,

large-impact bridge collapses. This results in tweets

that are pointing to the same incident, but are

mapped to different geographical cells. Users, for

example, tweet about events that are far away. A

differentiation, therefore, must be made between

ground users and other users. While most relevant

for bridges, this difference in location proves to be

applicable across infrastructures. The search terms

and stop words used to detect bridge-specific

damage events are listed in Table 1.

Table 1: Search terms and stop words for bridge damage

events.

Search Terms Stop Words

bridge {collapse, damaged,

closure, closed, flooded,

accident}

friendship, reopened, re-

opened, pending, fish, bid,

awe, awesome, wheelchair

Social Sensor

Twitter

Filtering

Search terms

Stop words

Geolocation

Event detection

Bridges

Highways

Gas lines

Power

Assign to cell

IoTBD 2016 - International Conference on Internet of Things and Big Data

436

3.2 Highways

Analysis of highway-related events is dependent on

the severity of the event considered. For example, it

was found that many Twitter users use the platform

to complain about delays and increased traffic times

on the highway, rather than to indicate actual

infrastructure damage. Considering only major

traffic or accidents that occur on the highway

decreases the amount of noise in the data. As many

highway damage events are due to natural disasters,

future filtering of the data in conjunction with

information on natural disasters may also decrease

noise and enable better detection of highway damage

events. The search terms and stop words used to

detect highway-related damage events are listed in

Table 2.

Table 2: Search terms and stop words for highway damage

events.

Search Terms Stop Words

highway {damaged, closed,

blocked, accident, mud,

pothole, snow, gridlock}

boating, watch, explore,

delays, symbolic

3.3 Gas Lines

The social sensor data filtered to detect gas line

damage events was the noisiest dataset of the

infrastructures studied. While the bridge dataset

includes differences in location between the tweet

and event of interest, the gas line dataset also

includes differences in time between the tweet and

event of interest. For example, users tweet about gas

leaks that have occurred in the past rather than about

the current state of the infrastructure. In addition,

irrelevant tweets include those complaining about

the smell of gas from cars at drive-throughs, or

about suspected but unsubstantiated gas leaks. Real

gas leaks or damage to gas lines can result in severe

health and safety consequences, so it is important to

be able to detect these events. The search terms and

stop words used to detect damage events related to

gas lines are listed in Table 3. Note that due to the

noise in this dataset and the number of stop words

needed to filter out irrelevant data, a representative

sample of, but not all, stop words are listed.

Table 3: Search terms and stop words for gas line damage

events.

Search Terms Stop Words

gas {leak, line}

plumbers, suspected, in-home estimate,

repairs underway, drive-through, drive-

thru, short line, tanker, contained, fixed

3.4 Power

In the data filtering process for power infrastructure,

we are able to detect both larger-scale power outages

that occur across cities and countries, e.g., the major

outage in Puerto Rico on October 23, 2015, as well

as smaller-scale individual outages, e.g., an outage

associated with a local elementary school. For the

stop words filter, we found that tweets containing

any permutation of two or more of the hashtags

#power, #outage, #blackout, or #grid were

irrelevant. This is due to the general meanings of

these words and the prevalence of these hashtags in

referring to things outside the scope of events of

interest. Over time, as different events occur and

memes develop that utilize words associated with

these critical public infrastructures but are unrelated

to actual infrastructure damage, the data filtering

system must be able to filter out this noise. In

addition, tweets relating to news stories of past

power outages, rather than the current state of power

infrastructure, have to be filtered out. Future filtering

in conjunction with text mining of news links in

articles will facilitate this filtering. The search terms

and stop words used to detect failure events of

power infrastructure are listed in Table 4.

Table 4: Search terms and stop words for power

infrastructure damage events.

Search Terms Stop Words

power outage

#power, #outage, #blackout, #grid,

back on, claims, resolved, files,

stories, hotel

4 EVALUATION OF APPROACH

In this section, we discuss the filtering efficiency of

the proposed approach, and show how results can be

visualized to facilitate detection, identification, and

inference about critical public infrastructure damage

and failure events.

Table 5 shows the number of social media items

downloaded and filtered through each step of the

data filtering process. The total number of tweets

remaining after each step for the four infrastructures

is listed. In addition, for filter steps two and three,

the percentage of data remaining after that filter step

compared to the previous step is given. The relative

number of tweets across the four infrastructures is an

indicator of the relative prevalence of tweets related

to those systems among Twitter users.

The initial filter based on search terms includes

items both relevant and irrelevant to the infrastructu-

Detection of Damage and Failure Events of Critical Public Infrastructure using Social Sensor Big Data

437

Table 5: Filtering results: number and percentage of tweets remaining after each filter step for four infrastructures of

interest: bridges, highways, gas lines, and power.

Infrastructure

Filter based on search terms Filter based on stop words Filter based on geolocation

Number of tweets Number of tweets % remaining Number of tweets % remaining

Bridges 8436 8364 99.1% 1673 20.0%

Highways 5826 5817 99.8% 2368 40.7%

Gas lines 8709 8417 96.6% 2249 26.7%

Power 6648 6474 97.4% 1127 17.4%

re damage events of interest. The stop words filter

out irrelevant tweets. From the first search-term

filter to the second stop-word filter steps, we see that

there are surprisingly low levels of noise in the

social sensor data. The percentage of data remaining

after the stop-word filter, however, is not 100%.

This noise must be filtered out using stop words.

This is important to ensure the minimization of the

number of incorrect detections of infrastructure

damage events.

Detections of damage or failure of critical public

infrastructure have significant societal and economic

impacts. If, for example, crews are dispatched to

repair certain infrastructure, emergency responders

are distributed to particular locations, or

infrastructures are closed for safety based on this

information, it is important that there is a high

confidence in the inference about the infrastructure

damage states before acting. This has policy

implications for the accuracy of inference based on

social sensor data required to transition from the

data and event detections to public or community

actions.

From Table 5, we see that in going from the

second stop-word filter step to the third geolocation

filter step, the number of results filtered out due to

incorrect or insufficient geolocation information is

significant. This is due to the presence of irrelevant

tweets, as well as to the lack of geolocation

information to confirm relevance of a tweet to an

event of interest. This demonstrates the need to

augment the social sensor data with other data

sources, including physical sensor data, news

sources, and alternate social sensor information.

Doing so will reduce the loss of information and

increase the resolution of the relevant information in

the third filtering step. This integration across data

sources will also facilitate automation in detection of

infrastructure damage or failure events.

4.1 Data Visualization

In addition to the detection of an event, given the

spatial distribution of public infrastructure, it is

important to be able to visualize the damage or

failure events. Figures 2-4 show visualizations of

events of interest, including the geolocated relevant

tweets and detected events.

Figure 2 shows a cluster of relevant tweets and

detected events in the Johannesburg, South Africa,

area related to bridge damage. The number of

relevant tweets in a concentrated geographical area,

i.e., the number of tweets mapped to a cell, can be

used as a measure of the intensity of an event. In

Figure 2, we see the relevant tweets detecting the

severe bridge collapse in Johannesburg on October

14, 2015. The distribution of tweets to different cells

is due to differences in identifications of

geolocations. In this case, geolocations for tweets

relevant to this event include Johannesburg,

Sandton, and Grayston Bridge. This is because the

bridge collapse event occurred in Johannesburg’s

Sandton district near Grayston Drive. Therefore,

tweets related to the same event can be mapped to

different cells due to different geolocations. Despite

the distribution across cells, the number of relevant

tweets in nearby cells indicates a severe event. In

this case, there were two deaths and 20 injured as a

result of this failure event.

Figure 2: Relevant tweets and detected bridge damage

events; example for Johannesburg, South Africa.

In Figure 3, we show an example of highway

damage-related relevant tweets and detected events

for California, USA. The figure shows the

correspondence between filtered, geolocated

relevant tweets and detected events. We are able to

detect damage events in both densely populated

urban areas, e.g., events in the San Francisco Bay

IoTBD 2016 - International Conference on Internet of Things and Big Data

438

Area, as well as in more sparsely populated rural

areas, e.g., events near Lone Mountain and Death

Valley. In addition, these results include a highway

damage event due to a flood and subsequent

mudslide, showing the ability of the approach to

detect damage events due to multiple hazards.

Figure 3: Relevant tweets and detected highway damage

events; example for California, USA.

For gas line damage, there were no particular

events of interest, so a map is not shown here. Maps

can, in general, be generated for locations or events

of interest. Power infrastructure damage events are

shown in Figure 4, which illustrates the widespread

nature of power failures. An example of relevant

tweets and detected events in the United States and

Caribbean are shown. In addition to the outage

events detected across the United States, we see the

major power outage detected in Puerto Rico from

the October 23, 2015, event.

Figure 4: Relevant tweets and detected power

infrastructure damage events; example for the United

States and Caribbean.

In general, we are able to use the social sensor

information to detect damage and failure events of

public infrastructure globally. The results are not

limited to any one country or region of the world, or

to the type or size of a community. Of course, event

detection relies on the presence of the social sensors,

e.g., Twitter data streams, but as social media

adoption increases around the world, the amount of

relevant data available will only increase.

5 RELATED WORK

The approach for public infrastructure damage and

failure event detection as described in this paper is

based on the LITMUS framework for landslide

detection built by Musaev et al., (2014a and 2014b).

A process similar to the LITMUS filtering process

was utilized to filter the noise out of infrastructure

damage-related tweets. However, this work differs

from the LITMUS work in that instead of detecting a

single type of event, we are focusing on different

infrastructures that can be damaged due to a variety

of events. For example, instead of detecting a

landslide, we are detecting damage to a highway that

may have been caused by a landslide or other event.

There have been several studies using social

sensor data to detect disaster events. This includes

studies related both to man-made hazards, e.g., mass

shootings (Vieweg et al., 2008; Palen et al., 2009);

and to natural hazards, e.g., earthquakes (Guy et al.,

2010; Sakaki et al., 2010; Caragea et al., 2011), fires

(Sutton et al., 2008), floods (Vieweg et al., 2010),

and tornadoes (Imran et al., 2013). Our work differs

from the disaster detection literature in that rather

than detection of widespread disaster events, we

detect damage to specific infrastructures, which may

or may not be related to or a result of a larger

disaster. In addition, many studies on detecting

disasters using social media data focus on the

detection or description of single hazards, whereas

the infrastructure damage events that we are looking

at may be caused by multiple hazards.

6 CONCLUSION

Detection of damage and failure events to public

infrastructure is critical to the ability of communities

around the world to minimize the risks associated

with both natural and man-made disasters and to

recover more quickly and efficiently from the

negative effects of these hazards. As many of our

public infrastructure systems are not physically

monitored to the degree necessary to provide

relevant, detailed information about the states of

these systems in real time, social sensor data is used

to perform this assessment and detect damage

events.

In this paper, we describe an approach to use

social sensor big data to identify public

Detection of Damage and Failure Events of Critical Public Infrastructure using Social Sensor Big Data

439

infrastructure damage events. This includes a three-

step filtering approach, whereby data is first filtered

using search terms relevant to the event of interest.

Next, noise in the data is filtered out using an

exclusion rule based on the presence of stop words.

Finally, data is filtered based on geolocation,

resulting in each relevant filtered data item being

assigned to a 2.5-minute by 2.5-minute cell in a grid

mapped to the surface of the Earth.

Once all relevant data are mapped to their

respective cells, all data in a single cell are assessed

to identify the infrastructure damage and failure

events. In this paper, we present results for detection

of damage events for transportation and energy

systems, and in particular for bridges, highways, gas

lines, and power infrastructure. We evaluate the

approach using real-world data collected from

October 2015. We show the ability of our approach

to use social sensor information, in this case Twitter

data streams, to detect damage events. In addition,

we show how results can be visualized to facilitate

detection, identification, and inference about

infrastructure damage.

As infrastructures are subjected to an increasing

number of hazards, the ability to detect and localize

damage events to these infrastructures is becoming

an increasingly important task to improve the

resilience of communities. In this paper, we

demonstrate the ability of and value in using social

sensor big data to detect damage and failure events

in these critical public infrastructures.

ACKNOWLEDGEMENTS

This work was partially funded by the National

Science Foundation through the CNS/CISE program

(Award #1541074). Any opinions, findings, and

conclusions or recommendations expressed in this

material are those of the authors and do not

necessarily reflect the views of the National Science

Foundation.

REFERENCES

Caragea, C., McNeese, N., Jaiswal, A., Traylor, G., Kim,

H., Mitra, P., Wu, D., Tapia, A.H., Giles, L., Jansen,

B.J., Yen, J., 2011. Classifying text messages for the

Haiti earthquake. In ISCRAM ’11, Lisbon, Portugal.

Cheng, Z., Caverlee, J., Lee, K., 2010. You are where you

tweet: A content-based approach to geo-locating

Twitter users. In CIKM’10, Toronto, Canada.

Google, https://developers.google.com/maps/documenta-

tion/geocoding/intro, accessed on 2/5/2016.

Guy, M., Earle, P., Ostrum, C., Gruchalla, K., Horvath, S.,

2010. Integration and dissemination of citizen reported

and seismically derived earthquake information via

social network technologies. In IDA’10, Tuscon,

Arizona.

Hecht, B., Hong, L., Suh, B., Chi, E.H., 2011. Tweets

from Justin Bieber’s heart: The dynamics of the

“location” field in user profiles. In CHI ’11,

Vancouver, Canada.

Imran, M., Elbassuoni, S., Castillo, C., Diaz, F., Meier, P.,

2013. Extracting information nuggets from disaster-

related messages in social media. In ISCRAM ’13,

Baden-Baden, Germany.

Manning, C.D., Surdeanu, M., Bauer, J., Finkel, J.,

Bethard, S.J., and McClosky, D., 2014. The Stanford

CoreNLP Natural Language Processing Toolkit.

Proceedings of the 52

Annual Meeting of the

Association for Computational Linguistics: System

Demonstrations, pp. 55-60, Baltimore, Maryland.

Musaev, A., Wang., D., Pu, C., 2014a. LITMUS:

Landslide detection by integration multiple sources. In

ISCRAM ’14, University Park, Pennsylvania.

Musaev, A., Wang., D., Cho, C.A., Pu., C., 2014b.

Landslide detection service based on composition of

physical and social information services. In ICWS ’14,

Anchorage, Alaska.

Palen, L., Vieweg, S., Liu, S., Hughes, A., 2009. Crisis in

a networked world: Features of computer-mediated

communication in the April 16, 2007 Virginia Tech

event. Social Science Computer Review Special Issue

on E-Social Science.

Sakaki, T., Okazaki, M., Matsuo, Y., 2010. Earthquake

shakes Twitter users: real-time event detection by

social sensors. In WWW ’10, Raleigh, North Carolina.

Sutton, J., Palen, L., Sklovaski, I., 2008. Backchannels on

the front lines: Emergent use of social media in the

2007 Southern California fires. In ISCRAM ’08,

Washington, DC.

Vieweg, S., Palen, L., Liu, S., Hughes, A., Sutton, J.,

2008. Collective intelligence in disaster: Examination

of the phenomenon in the aftermath of the 2007

Virginia Tech shooting. In ISCRAM ’08, Washington,

DC.

Vieweg, S., Hughes, A.L., Starbird, K., Palen, L., 2010.

Microblogging during two natural hazards events:

What Twitter may contribute to situational awareness.

In CHI ’10, Atlanta, Georgia.

IoTBD 2016 - International Conference on Internet of Things and Big Data

440