Professional Documents
Culture Documents
Abstract—In the aftermath of a natural disaster reconnaissance network of high accuracy accelerometers is needed which is
is typically conducted by teams of engineers tasked with tagging not practical with respect to cost or time of deployment.
buildings according to their damage state. Tagging (red, yellow, or With the advent of the Internet of things (IoT) era, a dense
green) conveys information about the condition of the building
(unsafe, needs further evaluation, or safe, respectively). While network of citizen owned devices is already deployed, such
thorough, the process can take several days to weeks to be as smart-phones, tablets, ... etc. These devices are already
completed. Automated assessment is an attractive alternative to equipped with vibration sensors, specifically accelerometers.
manual inspection but requires deploying a dense network of The objective of this work is to develop a framework that can
sensors at the granularity of each structure. Such a network was crowd-source relatively low-quality readings from distributed
deemed to be impractical with respect to cost or deployment
time. However, with the advent of the Internet of things (IoT) smart citizen owned devices and distill that information into
era, a massive network of citizen-owned smart devices such as actionable information. In prior work, an EEW system (called
tablets and smart-phones that contain vibration sensors (e.g. MyShake) was developed using only mobile sensors data [5].
accelerometers) is already deployed. The objective of this work MyShake successfully provided a 20 seconds warning for a
is to develop a framework that can crowd-source relatively low- magnitude 5 (M5) earthquake. Furthermore, CSN, another
quality readings from distributed smart citizen owned devices
and distill that information into actionable information. This earthquake detection system used phone accelerometers to
information can be provided to public safety personnel within capture the s-wave in addition to the much harder to detect
minutes of an event, in the form of a disaster map, with buildings p-wave [6]. While prior work focused on providing early
tagged by their most likely damage state. This paper reports on warning, we focus on utilizing mobile accelerometer reading
the development of an application running on a mobile phone to directly assess the structural health of buildings post event.
to collect readings, and coupled to a cloud based server, used to
generate the necessary tags. According to the Federal Emergency Management Agency
Index Terms—earthquakes; structural health monitoring; sen- (FEMA) the relative displacement between adjacent floors is
sors; smart-phones. related to building health [7] via the Inter-story Drift Ratios
(IDRs), which is calculated by (1). IDRs can be used to
I. I NTRODUCTION classify structures as immediate occupancy (IO), life safety
(LS), or collapse prevention (CP) depending on the maximum
Every year, earthquakes hit largely-populated cities destroy- instantaneous value of the IDR.
ing infrastructure and killing hundreds to thousands of people
Disp.F loori − Disp.F loori+1
depending on the severity of the event [1]. Recently, significant IDR = max (1)
F loorHeight
research attention has been devoted to studying the feasibility
of applying automated systems in detecting and reporting In this work, we use collected accelerometer readings to
seismic events. The two areas that are most prominently calculate the IDR for structures in the targeted area in order to
studied are Earthquake Early Warning (EEW) systems and infer their state. The rest of the paper is organized as follows.
post-disaster structural health assessment [2]. Several recent Section II provides an overview on the system architecture.
efforts targeted automatic structural health monitoring by using Section III reviews the experimental validation of the system.
satellite imaging [3], or by installing high accurate vibration Finally, the conclusion is drawn in section IV.
sensors [4]. Pros and cons of each inspection method are
summarized in Table I. The main disadvantage in adopting II. S YSTEM A RCHITECTURE
satellite monitoring is that it is capable of only capturing The system has been developed using client-server archi-
partial to complete building collapse. On the other hand, high tecture. The user application is used to detect, store and then
accuracy accelerometers are capable of detecting the building send recorded sensor data to a centralized cloud server. The
response to the event, from which its state can be inferred. server sorts, organizes and stores data in a database. When
However, to achieve large scale deployment, installing a dense requested by the user, the server computes the IDR value for
each building and classifies the building to IO, LS or CP. Then, 3) NTP-Time Synchronization: Calculating relative dis-
the server generates a map that displays buildings tagged with placements (IDR) requires millisecond accuracy for precision
their status. Fig. 1 shows the system block diagram. phasing. Therefore, a synchronization technique is required
across phones to avoid clock drift. the proposed application
A. Client Phone Application
uses Network Time Protocol (NTP) timing to ensure an
As part of this work, we developed a mobile-phone applica- accurate simultaneous reference for all devices [10].
tion that detects an earthquake and sends sensor readings to a
centralized cloud computing server along with the exact time, B. EC2 Cloud Server
device ID, user email address, building street address, and floor Using a cloud server has major advantages over using a
number. The first time the app is installed, the user is asked to standard server in terms of scalability, cost and avilability
manually enter an email address and the location information Amazon Web Services (AWS) is one of the major cloud
to be saved in a local database for future use. When the app computing service providers which make it suitable for the
is restarted, the stored information will be displayed and the proposed system. An Amazon Elastic Compute Cloud (EC2)
user will have the option to update it. A screen-shot of the is used as a base for all computations and processes needed
iOS application is in Fig. 2. which are illustrated below:
The app consists of 3 modes: steady mode, trigger mode, 1) MySQL Database: The system uses the MySQL
and streaming mode. The app has to enter a steady state database because of its ability to store and organize thousands
before it can be active and available to record an earthquake of readings to be recalled when needed. The data is organized
to avoid any additional noise in the reading. The app enters by zip code, street address, and floor number. The system
the steady mode when the absolute sensed acceleration in the consists of 3 different databases. One database is used to
x-y direction is below a certain threshold for a certain time store raw sensor data received from the app along with the
window. Trigger stage starts once an earthquake is detected, location information. Another database is used to store the
i.e. the predetermined threshold (0.1g) has been crossed in the resultant IDRs for each building along with the exact time and
x-y direction while the device is laid down horizontally. This date of the incident (earthquake). The last database is used as
technique was used in iShake [8] for earthquake detection. checkpoints for system debugging.
The app stores readings for a specific programmable duration. 2) Web page: A list of the tested buildings along with their
After that, the recorded information is sent to the centralized classification status is displayed on the system’s website which
cloud-server. An overview of how the app works is shown in is also hosted by the EC2 server. In addition, the website
Fig. 1. Key features of the application are: includes a map of the buildings tagged with their structural
1) Sampling Rate: The sampling rate of the accelerometer health state thus making it easier for the public to check their
directly affects displacement calculation. The error decreases buildings in the aftermath of an event. A screen-shot of the
with the increase of the sampling frequency. According to web page is shown in Fig. 3.
cellular phone manufacturers documentation, the maximum
frequency at which you can request updates is hardware- C. Classification Process
dependent [9]. In our experiments, the available phones were In the region of interest, the system loops over all the
iPhone 5 and 5S, where 100 Hz sampling rate is the highest buildings that have active users during the disaster to cate-
available option. gorize them into IO, LS or CP using FEMA standard [7]. For
2) Pre-Trigger: Pretrigger capability was added to reduce instance, an IDR over 5% for steel-moment frame buildings is
overall integration error. For that reason, a moving buffer was considered CP. IDRs over 2.5% and below 5% are classified as
added to store accelerometer data for 30 seconds. The stored LS, while IDRs less than 2.5% stated as being IO. Classifying
data is then sent to the server when the app is triggered. The a building is done through several stages. These stages are:
first 25 seconds of the pre-trigger data could be used for noise 1) Noise Cancellation: Since earthquake vibration starts
characterization and bias cancellation, and the last 5 seconds and ends at zero velocity, we used zero velocity update (ZUPT)
contain actual movement data. The error in the displacement technique for noise cancellation. Briefly, when the end of
without adding pre-trigger data compared with including pre- earthquake is detected the true velocity at this instant is exact
trigger data is discussed in the section III. zero, however, the measured one reflects the accumulation of
2019 IEEE Topical Conference on Wireless Sensors and Sensor Networks (WiSNet)
Raw
Start the app readings Request raw data from the database
database
Filter out accelerometer bias
NO Steady state
conditions Loop over floors, align data using NTP reference time
satisfied? Classification
YES
Status
database NO
Store last 30 second of readings as a buffer Are readings for
Go to the next floor
the same event?
Earthquake YES
NO Shaking?
Loop over buildings, align data and calculate the IDR
YES between every adjacent floors
Store accelerometer readings along with NTP exact timing
Classify buildings to IO, LS or CP using FEMA datasheet
Send the earthquake complete information including
sensors data, NTP timing, full address and floor number Store
Fig. 1: System Flowchart. The left hand side shows the mobile-phone application modes. The right hand side describes the
signal processing of received phone acceleration signals.
Fig. 2: Screenshots of the Client Application. The seismic Fig. 3: Website Screenshot. A detailed map is shown contain-
sensing is performed silently in the background. ing buildings tagged with their most likely structural health.
accelerometer noise in the preceding samples. With that infor- is used as the common time reference across device’s clock
mation, the error in displacement estimation during vibration is irrelevant. The experiments illustrated in section III shows
is minimized. The algorithm details are illustrated in [11]. the error reduction achieved by using NTP.
2) Displacement Calculation: Finding velocity as a func-
4) Relative Displacement Computation: The server loops
tion of time v(t) from raw R accelerometer data a(t) is done over adjacent floors in every building in the region of interest
using the equation v(t) = a(t) dt. The initial velocity v0 = 0
containing active devices during an earthquake. Then, the
because of the steady mode condition. Furthermore,R displace- maximum instantaneous IDR is computed using equation (2).
ment d(t) is found using the equation d(t) = v(t) dt. In
short, integrating the acceleration results in velocity while
d(t)i − d(t)i+1
double integrating the acceleration calculates displacement. IDRi = max (2)
h
Integration is done numerically by summation in the case of
digital (discrete) signals. Where IDRi is the maximum IDR value between floor i and
3) Data Synchronization: Data must be aligned before floor i + 1, d(t)i and d(t)i+1 are the calculated displacements
relative displacement calculation. NTP time is sent from the in meters for floor i and floor i + 1, respectively, and h is the
app with the acceleration data. As mentioned in section II-A3, floor height in meters (typically, 4m for steel-moment frame
2019 IEEE Topical Conference on Wireless Sensors and Sensor Networks (WiSNet)