QuakeWake: A Novel AI-Based Early Earthquake Warning and

Post-Quake Building Safety Guidance System

Dhroov V Bharatia

ACM Student Member, U.S.A.

Keywords:

Earthquake, Earthquake Early Warning, Supervised Learning, Autonomous Learning, P-Wave, S-Wave,

Building Damage, Dynamic Time Warping, User Collaboration.

Abstract:

Millions of people around the world suffer from earthquakes every year. This research introduces an inno-

vative, mobile device-based approach for real-time earthquake detection and prediction. By discerning quake

patterns from users’ regular usage patterns, a novel approach that prevents excessively draining battery uses an

on-device neural network only when needed to detect earthquake tremors. Cloud servers running an AI module

reliably predict the quake intensity and propagation pattern using signals from many users, enabling warning

others who have yet to experience these tremors. It also detects buildings at high risk to reinhabit due to high

relative ﬂoor displacement exceeding the building safety standards. A low-cost, affordable, and highly reliable

optional adjunct device on the user’s premise captures tremors with higher accuracy than mobile devices. This

enables effective building-wide earthquake warnings and eliminates fatalities due to post-earthquake build-

ing structural integrity issues. With a neural network trained with many past earthquake patterns, the mobile

devices reliably detected quakes and the AI module accurately detected its propagation with 99% accuracy,

warning users along its path. Moreover, the adjunct device adequately captured shifts in the building’s struc-

ture and reliably ﬂagged the building as uninhabitable with more than 95% accuracy.

1 INTRODUCTION

Earthquakes are very dangerous as they cause tremen-

dous damage and often there is not enough warning.

Earthquake Early Warning (EEW) systems attempt to

warn people to evacuate during an earthquake. How-

ever, earthquakes can strike at any time and at any

place and it is not always possible to accurately pre-

dict or detect earthquakes. Today’s earthquake detec-

tion methods using seismometers are too expensive,

too few, and not always reliable and the current meth-

ods have proven to be inadequate as apparent from the

many casualties every year. A system that can imme-

diately and reliably warn a large number of people to

evacuate during an earthquake can save lives.

Typical earthquakes emit two very common types

of waves: Primary (P)-Waves and Secondary (S)-

Waves. P-waves travel fast but are not very harmful

whereas S-waves travel about half the speed, but are

very harmful. Many EEW systems use sensors that

detect the P-Waves and warn people before the harm-

ful earthquake surface waves arrive. However, there

https://orcid.org/0009-0003-1502-8081

are too few detection systems to immediately and reli-

ably detect these P-Waves and warn a very large num-

ber of users in harm’s way.

The purpose of this research is to create an intelli-

gent earthquake system to reliably warn people before

an earthquake arrives and if their building is unsafe af-

ter the earthquake. A novel collaborative earthquake

early warning system is presented using the ubiqui-

tous mobile smartphones. This helps overcome the

two big issues faced by conventional EEW systems -

the number of sensors providing coverage over popu-

lous areas and reliable means to notify affected users

when an earthquake is detected. The accelerometer in

a mobile device monitors the user’s phone movements

and a pre-trained neural network on the device helps

determine if such movements may be due to an earth-

quake. Collaborating such detection across a large

number of users in the vicinity adds accuracy and re-

liability to such detection. Additionally, by predicting

the earthquake’s path of progression, users in the path

of the earthquake can also be warned even if they have

not yet experienced any tremors. The ground shift and

intensity recorded by each device further help in es-

timating if their building is safe to inhibit and hence

Bharatia, D. V.

QuakeWake: A Novel AI-Based Early Earthquake Warning and Post-Quake Building Safety Guidance System.

DOI: 10.5220/0013123200003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 303-310

ISBN: 978-989-758-737-5; ISSN: 2184-433X

303

provides additional valuable safety guidance during

and after an earthquake. Additionally, low-cost ex-

ternal accelerometer sensors mounted on the building

structure can further ensure in providing such reliable

building safety guidance.

1.1 Related Work

In 1868, the earliest front-detection EEW system

(Cooper, 1868) was for the Hayward fault near San

Francisco where telegraph cables signaled ground

shaking to ring the city bell. (Aranda et al., 1995)

more recently used front-detection to give 60s warn-

ing for Mexico City 320km away. But awaiting strong

ground shaking loses valuable time and directional

propagation prevents warnings for large areas around

the epicenter leading to P-Waves based detection.

Reliably detecting P-Waves and using its ampli-

tude to judge quake intensity poses big challenges due

to limited detection time constraints as apparent in

(Allen and Kanamori, 2003), (Allen et al., 2009) and

(Wu and Kanamori, 2008). Most of EEW work still

focuses on seismometers or a network of seismome-

ters leading to two general categories of modern EEW

systems: 1) regional and 2) on-site EEW.

Regional EEWs use a few seconds of initial

ground motion data to provide intensity measure es-

timates over a region using pre-calibrated ground-

motion models. As in (Bhardwaj et al., 2016), (Mu

and Yuen, 2016), and (M

unchmeyer et al., 2021) they

face signiﬁcant issues due to epistemic and aleatory

uncertainties in modeling and epicenter proximity

constraints. As in (Caruso et al., 2017), (M

unchmeyer

et al., 2021) and (Colombelli and Zollo, 2015) on-site

EEWs focus on standalone single sensor estimating

intensity and surface waves based on early P-Waves.

They are limited to a locality and fail to help address

structures in the area.

Building damage due to earthquakes involves a

qualiﬁed inspector and since most buildings lack ﬂoor

shift detection it prevents automated warnings. Re-

cent machine learning techniques can detect certain

cracks in structure such as in (Chen et al., 2024) but

hidden building damage remains an unresolved issue.

(Reilly et al., 2013) used mobile device ac-

celerometers to detect ground motion and transfer the

accelerometer samples to a server for ground motion

analysis. However, such continuous sever-based anal-

ysis is not pragmatic on a large scale. Instead, by

having a very large number of user phones, acting

as earthquake monitoring sensors, this paper demon-

strates how to achieve very high reliability, accuracy,

and coverage for earthquake detection without the

aforesaid limitations of conventional EEW systems.

PiCtrl Server

Raspberry Pi Zero W

MPU6050

Accelerometer

I2C

Audio

Warning

WiFi

Android Device

QuakeWake Room

Database

QuakeWake

Service

Operation Manager

QuakeWake

Activities

Notifier

PiManager

Normal Operation

Calibration

Provisioning

Accelerometer

Quakewake

Cloud

Internet

Figure 1: QuakeWake Overview.

2 PROPOSED APPROACH

QuakeWake leverages the pervasiveness of mobile

phones equipped with an accelerometer to detect and

collaborate earthquake tremors along with their loca-

tion as shown in Figure 1. It reliably projects the path,

intensity estimate, and time of the earthquake’s arrival

even for numerous users not yet in the earthquake’s

vicinity. Additionally, using the relative ﬂoor-shift

data of a user’s building gathered during the earth-

quake, it warns if the building is safe to reinhabit.

A mobile device is typically equipped with

a micro-electro-mechanical system (MEMS) ac-

celerometer that measures the acceleration in move-

ment along the X, Y, and Z axis allowing analysis of

the device’s movement. Unlike other user activity-

related movements, earthquake tremors caused by

fast-moving P-waves produce a distinct pattern of this

movement. A pre-trained convolution neural network

(CNN) as in Figure 3 reliably detects this movement

and warns users of the earthquake. The speed dif-

ference between these early detected P-waves and

slow-moving harmful surface waves like S-waves en-

abled early warning. By extrapolating the location of

detected tremors from many users, QuakeWake also

warns users not yet in the vicinity, giving ample time

to evacuate.

A building structure that experiences a large shift

is prone to severe structural damage that could be

hidden from view and extremely unsafe to inhabit.

QuakeWake detects such shifts during the earthquake

and based on collaborated data from similar buildings

in the neighborhood, it uses a cloud-based neural net-

work to predict building safety to warn users. An

optional external accelerometer device powered by a

low-cost Raspberry Pi Zero W can be mounted on the

user’s premise to further improve this building safety

guidance. It improves reliability by providing better

relative displacement detection and its low cost makes

it a pragmatic user safety feature for all buildings.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

304

2000 4000 6000 8000 10000

Distance from focus (km)

P-Wave

S-Wave

Travel Time

(minutes)

P to S

Time Interval

Figure 2: Wave Lag.

• Input: 1000 timesteps, 3 features.

• Conv1D: filters 128, kernel-size: 3,

ReLU

• Conv1D: filters 128, kernel-size: 3,

• Dropout(0.5)

• MaxPooling1D: pool_size: 2

• Flatten

• Dense: 256, activation: ReLU

• Dense: 5, activation: softmax

• Loss: category_crossentropy, optimizer=adam,

metrics=accuracy

ReLU

Conv1D

Dense

None

1-D

Figure 3: Detect-CNN.

2.1 Earthquake Waves

When an earthquake occurs, the shockwaves of re-

leased energy take the form of seismic waves. Seismic

waves are categorized as primary waves (P-Wave),

secondary waves (S-Wave), and additional surface

waves (L-Wave/R-Wave).

Primary waves (P-Waves) are the least destructive

type of seismic wave and have the fastest velocity.

They are longitudinal waves that travel fastest at 6

km/s, pass through all matter, and cause objects to

shift back and forth. P-Waves being the fastest ar-

rive early and are used to detect earthquakes before

the more destructive waves hit. As shown in ﬁgure

2, the distance of the user from the epicenter deter-

mines the fore-warning duration based on how early

P-Waves arrive before other waves.

Secondary waves (S-Waves) are the more destruc-

tive large amplitude waves generated by tectonic plate

movements. Other major destructive wave types in-

clude Love (L-Waves) and Rayleigh (R-Waves).

2.2 Detecting Earthquake

Earthquake causes prolonged shaking of the earth’s

surface proportional to the quake’s intensity and this

causes a phone resting on a surface to move with it.

The acceleration of this movement is monitored by

the phone’s accelerometer and by analyzing the shake

pattern, an earthquake is detected. Earthquakes can

occur at any time, and there are many wave patterns

depending on the type of waves and the user’s relative

position to the quake epicenter. Therefore, the Quake-

Wake app is provisioned to continuously monitor for

tremors. Additionally, to rule out other shaking mo-

tions such as user movements, notiﬁcations with hap-

tic vibrations, and other causes of seismic waves, it is

essential to reliably detect an earthquake pattern. Di-

rect point-wise comparison of waveforms fails to reli-

ably discern shake patterns, especially for low quake

intensity P-Waves. Therefore, a 1-dimensional (1-D)

CNN is used to accurately detect the earthquake shake

patterns.

The accelerometer motion data is sampled with a

sliding window for motion along the X, Y, and Z axis.

Overlapping 10s waves are sampled at 100Hz provid-

ing 1000 samples, representing a feature per channel

which is fed as input to the CNN and the output pre-

dicts if a quake is detected. This detection CNN was

trained using publicly available earthquake datasets

such as STEAD (Mousavi et al., 2019). If an optional

external accelerometer device is present, QuakeWake

app also validates its results with it.

2.3 Activity Matching with Dynamic

Time Warping

Frequent predictions using detection CNN are re-

source intensive and heavily drain the phone battery.

To conserve battery, it is necessary to attempt such

prediction only when there is something unusual as

there is no earthquake most of the time.

Users tend to repeat a ﬁnite set A of common daily

activities. A time-series t

of an activity typically does

not exactly match the time-series t

for a known activ-

ity y ∈ A but they do loosely match barring some tem-

poral variations. So a simple point-wise Euclidean

distance cannot correlate t

and t

. As shown in Fig-

ure 4, an optimal alignment between t

and t

needs

non-linear “warping” (shown as green dotted lines be-

tween them) by stretching or shrinking along the time

axis. A Dynamic Time Warping (DTW) algorithm

non-linearly warp matches similar activities even if

their time-series are out of phase along the time axis.

As shown in (Salvador and Chan, 2007), DTW (X ,Y )

provides distances between the i

point in time-series

for waveform X to j

point in time-series t

for

waveform Y aligning along optimal path P as:

DTW (X ,Y ) =

∑

(i, j)∈P

−Y

)

(1)

DTW operates in linear time and space and can

quickly correlate a current activity to an everyday

known activity. Use of resource-intensive CNN can

be restricted to only most likely unusual earthquake

samples, signiﬁcantly reducing the battery drain.

A time-series sample s comprises a tuple

) with 0 ≤ i ≤ w

, and w

represents to-

tal intervals in s’s capture window. s

is the

acceleration along X,Y , Z axis for i

interval with

QuakeWake: A Novel AI-Based Early Earthquake Warning and Post-Quake Building Safety Guidance System

305

Figure 4: Dynamic Time Warping.

positive values when phone moves to right sideways:

X axis, away from user: Y axis, vertically up: Z

axis. At a 100Hz sampling rate, s of duration t has

= t × 100 intervals. Early detection time-series

spans m sec and earthquakes conﬁrming time-series

spans nsec,m < n. As per Figure 2, forewarning

needs at least m ≈ 2sec and n ≈ 10sec.

The system manages time-series activity samples

saved in three sets. An Earthquake Set E has m sec

samples of potential earthquake patterns e ∈ E. The

Activity Set A for activity a ∈ A, each with ﬁrst m sec

samples of a known user activity. The Probable-

activity Set P has msec samples of activity p ∈ P to be

potential user activity known not to be an earthquake.

A current sample is searched in a set using a novel

DTW exact indexing technique based on (Keogh and

Ratanamahatana, 2005). A query sample Q comprises

a partial incoming time-series which is extended fur-

ther only for high likelihood of a match. Consider

each existing set sample C is n units long and N is the

dimensionality of index space 1 ≤ N ≤ n. U and L

are upper and lower warping path envelopes around

query Q with

and

being piecewise constant ap-

proximations of U and L for time-series represented

as c

,1 ≤ i ≤ N. Lower bound piecewise aggregate

approximation LB PAA(Q,C) approximates the lower

bound distance of query Q from sample C in an exist-

ing set as:

∑

i=1







−

)

if c

> U

−

)

if c

> L

0 otherwise

(2)

LB PAA(Q,C) ≤ T

The MINDIST (Q, R) is the lower bounding mea-

sure between query Q and a minimum bounding rect-

angle (MBR) R that represents the smallest rectangle

to spatially contain piecewise aggregate approxima-

tion points C

under a node of a hierarchical MBR-

based indexing structure, with MINDIST(Q, R) as:

∑

i=1







−

)

if l

> U

−

)

if h

> L

0 otherwise

(3)

MINDIST (Q, R) ≤ T

When LB PAA(Q,C) exceeds T

there is no like-

lihood of match but when it is less, it is possible

foreach user awaiting earthquake detection do

s ← awaitMovement(): Monitor time-series s captured

with < m sec window around every m/8 sec until 2

consecutive samples s for each interval 0 ≤ i ≤ w

have

{

) | |s

| > T

,|s

| > T

or |s

| > T

}

, where

are minimum move thresholds.;

if s ∈ W // Known waking activity then

awaitCalm(): Wait for

{

) | |s

| < T

,|s

| < T

,|s

| < T

}

for

each sample interval 0 ≤ i ≤ w

;

continue;

end

if s ∈ E or detectCNN(s) := possible earthquake // Known

tremor or suspected tremor detected then

earthquakePossibleWarning();

con f irmCNN(): Capture more samples and form n sec

time-series p, with each s as m sec of detectCNN(s);

con f irm ← detectCNN(p);

if con f irm = earthquake then

earthquakeCon f irmedWarning();

inQuakeMonitorAndReport();

end

clearEarthquake();

continue;

end

if s ∈ P // Learn more waking activities then

Increment counter associated with s in P;

if counter(s ∈ P) > T

, where T

is activity inclusion

threshold then

Move s from P to A;

end

else

P ← P ∪ {s};

end

Algorithm 1: Earthquake Detection Algorithm.

that Q may match C and therefore additional sam-

ple intervals are retrieved if Q does yet have sufﬁcient

sample intervals as C. This technique is also appli-

cable while computing MINDIST (Q,R). It enables

handling samples with time-axis warping with one or

more time-series portions expanded beyond that of the

target time-series. This technique allows to start ag-

gressively indexing samples even when the full win-

dow of the query sample is not yet complete and once

it is known to be highly probable for matching, it can

be expanded to the full query time-series with mini-

mal cost. It avoids awaiting query to be fully available

saving valuable earthquake detection time.

Algorithm 1, detect earthquakes with very few

computations. awaitMovement() helps avoid com-

putation when the phone lies dormant as on a ta-

ble. When a known user activity disrupts this dor-

mant state, all samples during this activity are ig-

nored until the phone is dormant, eliminating pro-

cessing when the user is using the phone. In the dor-

mant state, only if a known potential earthquake pat-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

306

4. notifyQuake

(eta, intensity)

4. notifyQuake

(eta, intensity)

Quakewake

Cloud

3. Analyze

propagation

1. Wave Detected

2. quakeAt

(gps, intensity)

2. quakeAt

(gps, intensity)

Figure 5: Collaboration Protocol.

90 secs

190 secs

Extrapolations

Reported

140 secs

Epicenter

Line of Best Fit

Figure 6: Predict Path.

tern or detectCNN() detects a possible earthquake,

earthquake processing is launched. As it learns more

non-earthquake activities that wake it from a dormant

state, A expands, and it spends most of the time in

awaitMovement() and awaitCalm() - both requiring

very little computation, thereby conserving battery

power. Moreover, earthquake warnings can be issued

as early as m = 2 seconds which is a great improve-

ment over today’s EEW systems.

2.4 Collaborating Early Warning

A QuakeWake mobile application immediately warns

its user when it detects earthquake tremors. Addition-

ally, as shown in Figure 5, it also reports the detected

earthquake intensity and its GPS location to Quake-

Wake cloud servers. QuakeWake cloud then creates

an earthquake activity map and as shown in Figure

6, extrapolates the earthquake intensity, direction of

propagation, and estimated time of arrival (ETA) to

wider areas around the epicenter and notiﬁes users in

these affected areas giving them an additional 90 to

190 seconds of advance notiﬁcation, potentially sav-

ing their lives. Reports from many users in highly

populated urban areas signiﬁcantly increase predic-

tion accuracy for the path and ETA.

Quakewake

Cloud

2. Analysis

1. isSafe(gps, localIntensity)

3. safetyRecommend

(gps, expected-intensity, isSafe)

Figure 7: Building Safety.

2.5 Building Safety

Building safety standards and organizations such as

the United States Geological Survey (USGS, 2024)

specify a safe limit for ﬂoor movement relative to the

structure below. Due to their high cost, most buildings

are not equipped with the instrumentation that mea-

sures such a shift during an earthquake. The hidden

structural damage due to such shift is extremely dan-

gerous leaving the building residents vulnerable dur-

ing and after major earthquakes.

Although no two buildings suffer the same dam-

age, similar buildings in a neighborhood tend to ex-

perience similar ﬂoor shift patterns and tend to suf-

fer similar earthquake damage. Therefore, a building-

safety neural network is used to predict the quake in-

tensity and shifts using regression as shown in Figure

8 and warn users of unsafe buildings. It can guide

users even if they could not record the precise earth-

quake shift for their building. This neural network

is pre-trained using historical datasets for earthquake

damage and safety information, and continuously re-

ﬁned with current quake data using an adjunct net-

work model. Building structures respond differently

to different types of earthquake waves. Even for the

same wave type, response changes based on whether

the building was exposed to a rising or falling edge

of the surface wave. Different buildings on the same

path can be exposed to different cycles of the destruc-

tive S, R, or L waves. Therefore, the building-safety

neural network must be retrained with actual reports

from the affected buildings in the current earthquake

to discern these variations of destruction patterns in

the area. When a large number of users are concen-

trated in a small area, adequate training makes predic-

tion reliable. Sparsely populated areas may need ad-

ditional information such as ground shaking intensity

ShakeMap published by the U.S. Geological Survey.

QuakeWake also provides a means to use additional

building information such as from OpenStreetMap

(OSM) project (OpenStreetMap, 2024) based on the

user’s location and predict building safety as corre-

lated with other similar buildings in the neighbor-

hood.

The QuakeWake app records the building move-

ments during an earthquake. When a user requests a

building safety check, as shown in Figure 7 the app

sends a request to the QuakeWake cloud server with

the building type, building’s GPS location, recorded

intensity, and maximum local shifts recorded during

the quake. QuakeWake server uses a building-safety

neural network to predict the expected quake intensity

and whether it is safe.

QuakeWake: A Novel AI-Based Early Earthquake Warning and Post-Quake Building Safety Guidance System

307

-2

-1.5

-1

-0.5

0.5

1.5

-2

-1

0.2

0.4

0.6

0.8

0.2

0.4

0.6

0.8

Trained

Not-trained:

Neural Network uses regression

Geo-location

plane

Quake

Intensity

Trained Neural Network

Intensity Profile

Figure 8: Building Safety Regression.

3 RESULTS AND DISCUSSION

3.1 Experimental Setup

A set of randomized time-series samples representing

different activities and events was prepared and then

used to test different aspects of QuakeWake as dis-

cussed below.

Earthquake Samples. Time-series data from the

STEAD dataset as in (Mousavi et al., 2019) was cap-

tured along the three orthogonal seismograph com-

ponent axes. 5 sets of local earthquake samples

in randomized groups of 100,000 samples were ex-

tracted from the available earthquake samples world-

wide within 350km of the earthquakes. Each earth-

quake sample window contained both P and S waves

and began from 5 to 10 seconds prior to the P arrival

and ended at least 5 seconds after the S arrival. The

samples were conditioned by removing the mean and

resampled at 100 Hz. Similarly, a randomized group

of 100,000 non-earthquake seismic noise sample data

was also used from this dataset to represent seismic

conditions when there is no earthquake. Addition-

ally, the building damage dataset from the Gorkha

Earthquake of Nepal as in (M

obius, 2015) was used

to build 10 sets of building safety samples in random-

ized groups of 50,000 samples.

User Activity Samples. The Bogazici Univer-

sity smartphone accelerometer sensor dataset as in

(Davarcı and Anarım, 2022) was used in conjunction

with other real-time data collection by the researcher

for user movements during different activities. A cus-

tom app gathered time-series samples for different age

groups and genders by sampling accelerometer data

at 100Hz. 25 randomized groups of 10,000 samples

were created for typical everyday activities.

Motion Test-Bed. A set of modern Android phone

devices was used to test various aspects of Quake-

Wake. For accurate and repeatable test movements,

custom directional stepper motors from a 3D-printer

were repurposed to move a platform as per time-series

input data. The test setup was calibrated to produce

test device accelerometer data that matches the input

time-series sample to produce the platform motion.

Battery Drain. QuakeWake was tested with vary-

ing duty cycles of any-activity, each cycle compris-

ing a period of calm followed by a period of activ-

ity. The period of activity comprises a speciﬁc duty

cycle of earthquake activity, wherein a random mix

of earthquake and user activities are performed so as

to achieve the required earthquake duty cycle. Tests

were repeated with different any-activity duty cycles,

each repeated with different earthquake duty cycles

and intensities. The battery drain was recorded af-

ter an extended test duration depending on the any-

activity duty cycle.

Earthquake Detection. Earthquake detection was

tested by placing a test phone onto the motion test-

bed described above and driven by a test time-series

sample. Accuracy was determined based on whether

QuakeWake correctly detected earthquake samples

and did not detect other samples as earthquakes. Mul-

tiple trials were conducted with test samples for dif-

ferent earthquake shift distances, different earthquake

powers resulting in different peak acceleration magni-

tudes, different contact surfaces of the table, different

materials of the phone cover, and different earthquake

wave types. Tests were repeated continuously with

more than 98% user activity samples and the remain-

der as earthquake samples for a long time to ensure

that Dynamic Time Warping mostly eliminated the

resource-intensive detection CNN computations and

conserved battery power.

Training Building Safety Neural Network. The

building safety neural network was trained exten-

sively using samples from the Gorkha Earthquake

dataset. Testing for impact on building safety predic-

tion by retraining with reports from the current quake

was tested by retraining the preexisting trained model

with an additional set of 50,000 test samples such that

the test samples are localized in different high-density

reporting areas with at least one region each of 100,

50, 20, and 10 reports per square kilometer.

3.2 Comparisons and Analysis

As shown in Figures 9, 10, 11, and 12, QuakeWake

reliably detected an earthquake with very high accu-

racy for typical operating conditions. As shown in

Figures 9 and 10, with a phone on a very smooth con-

tact surface, it detected typical mild quakes with shifts

exceeding 40mm and peak acceleration of 10m/s

Figure 11 shows that for typical contact surfaces

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

308

Table 1: QuakeWake Test Summary.

Purpose Values

Detect varying shifts Mean:98.51% σ:0.96%

Detect varying power Mean:96.37% σ:2.96%

Detect surfaces Mean:89.01% σ:10.19%

Detect wave-types Mean:93.26% σ:5.89%

Collab. forewarning 78-187s Mean:150s

Bldg. safety accuracy Mean:90.43%, σ:2.09%

Bldg. retrain accuracy Mean:97.12% σ:1.26%

100

20 40 60 80 100 120 140

Accuracy (%)

Shift Distance (mm)

X-Accuracy

Y-Accuracy

Z-Accuracy

Figure 9: Detect varying shift.

of wooden furniture or with TPU, rubber, or sili-

cone covers, accuracy increased drastically resulting

in ﬂawless earthquake detections even for very mild

earthquakes. The typical contact surface does not re-

sult in excessive slippage and hence the ground mo-

tion during an earthquake tremor reaches the phone’s

accelerometer verbatim leading to a very high earth-

quake detection accuracy.

As evident from Figure 12, the earthquake detec-

tion accuracy is almost perfect for both the primary

P-waves and secondary S-waves but goes down by

around 15% for only the R-waves and by 8% when

only L-waves are used. As R-waves and L-waves are

not as common as the P-waves and S-waves, the neu-

ral network lacked training for L and R waves. This

is less of a concern as R and L-waves rarely exist on

their own and are always preceded by the P-waves.

Figure 13, shows results for any-activity duty cy-

cle of 50% revealing that the battery drain percent-

100

10 20 30 40 50

Accuracy (%)

Acceleration (m/s

)

X-Accuracy

Y-Accuracy

Z-Accuracy

Figure 10: Detect varying power.

100

WoodTPU WoodBare TileTPU TileBare

Accuracy (%)

Surface

X-Accuracy

Y-Accuracy

Z-Accuracy

Figure 11: Detect varying contact surfaces.

100

P-Wave

S-Wave

R-Wave L-Wave

Accuracy (%)

Earthquake Wave-Type

X-Accuracy

Y-Accuracy

Z-Accuracy

Figure 12: Detect varying wave-type.

age is mostly proportional to the duty cycle of the

earthquakes, conﬁrming that DTW-based activity de-

tection and use of CNN only for earthquake samples

works well. Severe earthquakes have less ambiguous

patterns leading to lower battery drain. Battery drain

for moderate earthquakes was not signiﬁcantly worse

especially when the duty cycle was low. Typically

earthquake duty cycles are low and QuakeWake pro-

vides excellent performance with a very low battery

drain allowing users to constantly use QuakeWake.

The baseline building safety neural network pre-

dicts with over 90% accuracy. As evident from Fig-

ure 14, when the earthquake passes through a densely

populated area, the safety accuracy prediction is near

perfect at 100 records/sq. km. and around 98% at 50

records/sq. km. As portrayed in Figure 8, more re-

training samples from small populated areas lead to

20 25 30 35

Battery Drain(%)

Earthquake Duty Cycle

Severe

Moderate

Figure 13: Battery Drain.

QuakeWake: A Novel AI-Based Early Earthquake Warning and Post-Quake Building Safety Guidance System

309

100

10 20 50 100

Safety Accuracy

Reports (per sq. km)

Figure 14: Building safety on retraining.

better regression and these retraining samples capture

the unique destruction pattern of the current earth-

quake leading to highly precise predictions.

4 CONCLUSIONS

QuakeWake reliably detects quake vibrations using

the accelerometer in phone devices and immediately

warns its users when it detects earthquake tremors or

if a user is in the path of an earthquake. Without

an elaborate infrastructure of costly seismographs, it

arms everyday users’ smartphones to detect an earth-

quake and forewarn users in harm’s way. It performs

well with typical non-slipping phone covers and on

wooden furniture. By using a novel Dynamic Time

Warping algorithm, it discerns everyday activity mo-

tions and uses the resource-intensive CNN detection

only when an earthquake is suspected, thereby con-

serving battery power. QuakeWake records the maxi-

mum shift experienced during an earthquake and uses

this information to enable building safety warnings.

By continuously retraining a building safety neural

network, it learns to predict this safety based on pat-

terns of the current earthquake. With its low-cost and

simple but robust design, QuakeWake can help save

numerous lives across the globe. In the future, Quake-

Wake can be integrated with other warning systems to

greatly enhance its efﬁcacy.

REFERENCES

Allen, R. M., Brown, H., Hellweg, M., Khainovski, O.,

Lombard, P., and Neuhauser, D. (2009). Real-time

earthquake detection and hazard assessment by elarms

across california. Geophysical research letters, 36(5).

Allen, R. M. and Kanamori, H. (2003). The potential for

earthquake early warning in southern california. Sci-

ence, 300(5620):786–789.

Aranda, J. E., Jimenez, A., Ibarrola, G., Alcantar, F.,

Aguilar, A., Inostroza, M., and Maldonado, S. (1995).

Mexico city seismic alert system. Seismological Re-

search Letters, 66(6):42–53.

Bhardwaj, R., Sharma, M. L., and Kumar, A. (2016). Multi-

parameter algorithm for earthquake early warning.

Geomatics, Natural Hazards and Risk, 7(4):1242–

1264.

Caruso, A., Colombelli, S., Elia, L., Picozzi, M., and Zollo,

A. (2017). An on-site alert level early warning sys-

tem for italy. Journal of Geophysical Research: Solid

Earth, 122(3):2106–2118.

Chen, W., Hou, J., Wang, Y., and Yu, M. (2024). Visu-

alization analysis of concrete crack detection in civil

engineering infrastructure based on knowledge graph.

Case Studies in Construction Materials, page e03711.

Colombelli, S. and Zollo, A. (2015). Fast determination

of earthquake magnitude and fault extent from real-

time p-wave recordings. Geophysical Journal Inter-

national, 202(2):1158–1163.

Cooper, J. (1868). Earthquake indicator. San Francisco

Bulletin, 3.

Davarcı, E. and Anarım, E. (2022). Bogazici university

smartphone accelerometer sensor dataset. Data in

brief, 40:107833.

Keogh, E. and Ratanamahatana, C. A. (2005). Exact index-

ing of dynamic time warping. Knowledge and infor-

mation systems, 7(3):358–386.

Mousavi, S. M., Sheng, Y., Zhu, W., and Beroza, G. C.

(2019). Stanford earthquake dataset (stead): A global

data set of seismic signals for ai. IEEE Access,

7:179464–179476.

Mu, H.-Q. and Yuen, K.-V. (2016). Ground motion predic-

tion equation development by heterogeneous bayesian

learning. Computer-Aided Civil and Infrastructure

Engineering, 31(10):761–776.

unchmeyer, J., Bindi, D., Leser, U., and Tilmann, F.

(2021). The transformer earthquake alerting model:

A new versatile approach to earthquake early warning.

Geophysical Journal International, 225(1):646–656.

obius (2015). Earthquake magnitude, dam-

age and impact: Gorkha earthquake 2015.

https://www.kaggle.com/datasets/arashnic/

earthquake-magnitude-damage-and-impact.

OpenStreetMap (2024). Openstreetmap project. https:

//www.openstreetmap.org.

Reilly, J., Dashti, S., Ervasti, M., Bray, J. D., Glaser, S. D.,

and Bayen, A. M. (2013). Mobile phones as seis-

mologic sensors: Automating data extraction for the

ishake system. IEEE Transactions on Automation Sci-

ence and Engineering, 10(2):242–251.

Salvador, S. and Chan, P. (2007). Toward accurate dynamic

time warping in linear time and space. Intelligent Data

Analysis, 11(5):561–580.

USGS (2024). United states geological survey scientiﬁc re-

search results. https://www.usgs.gov/products.

Wu, Y.-M. and Kanamori, H. (2008). Development of

an earthquake early warning system using real-time

strong motion signals. Sensors, 8(1):1–9.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

310