Professional Documents
Culture Documents
An Engineering Application of Earthquake Early Warning ePAD-Based Decision Framework For Elevator Control
An Engineering Application of Earthquake Early Warning ePAD-Based Decision Framework For Elevator Control
An Engineering Application of Earthquake Early Warning ePAD-Based Decision Framework For Elevator Control
Abstract: In a medium-to-large earthquake, there are often reports of people being trapped or injured in elevators. This study investigates
using an earthquake early warning (EEW) system, which provides seconds to tens of seconds warning before seismic waves arrive at a site,
to help people escape from the elevators before a strong shaking arrives. However, such an application remains as a major engineering
challenge due to the uncertainty of the EEW information and the short lead time available. A recent study presented an earthquake prob-
ability-based automated decision-making (ePAD) framework to address these issues. This paper focuses on studying the influence of two
commonly ignored factors, uncertainty of warning and lead time, on the decision of stopping the elevators and opening the doors when an
EEW message is received. Application of the ePAD framework requires using the performance-based earthquake engineering methodology
for elevator damage prediction, making decision based on a cost-benefit model and reducing computational time with a surrogate model.
The authors’ results show that ePAD can provide rational decisions for elevator control based on EEW information under different amounts
of lead time and uncertainty level of the warning. DOI: 10.1061/(ASCE)ST.1943-541X.0001356. © 2015 American Society of Civil
Engineers.
Author keywords: Performance-based earthquake engineering; Elevator control; Surrogate model; Earthquake early warning;
Decision-making; Safety and reliability.
between seismic network signals and seismic wave speeds presents a probability-based approach that utilizes data from the
(P-wave, S-wave, and surface waves) to estimate and broadcast rel- California Strong Motion Instrumentation Program database. A
evant earthquake information, such as the earthquake magnitude, threshold to stop the elevators is chosen based on a trade-off be-
hypocenter location, origin time, and local intensity measure tween the empirical false alarm and missed alarm rate. In this paper,
(Hilbring et al. 2014), before the arrival of damaging waves. the authors investigate the influence of the uncertainties in the EEW
Typical warning time ranges from a few seconds to a minute or information and warning lead time on decision making, which were
so, depending on various factors, such as seismic network density, not included in the previous studies.
geological structure of the region, and the site-to-hypocenter dis-
tance. Recently, an EEW system called the California integrated
seismic network (CISN) ShakeAlert, has been developed in ePAD Framework
California. The system combines the outputs of three distinct The ePAD framework (Wu et al. 2013) utilizes the EEW informa-
early warning algorithms: τ c − Pd on-site algorithm (Böse et al. tion to robustly and rapidly choose an optimal action from a set of
2009a, b), earthquake alarm systems (ElarmS) (Allen et al. possible mitigation actions Ω ¼ fa0 ; a1 ; : : : ; an g, where ai de-
2009a), and virtual seismologist (V-S) (Cua and Heaton 2007). notes initiating a certain action for i ¼ 1, : : : ,n, and a0 denotes
Fig. 1 demonstrates a sample user interface of the CISN ShakeAlert not to initiate any action. It makes decisions based on fundamental
system. Detailed information can be found in Böse et al. (2014). In decision theory (Raiffa and Schlaifer 1961; Von Neumann and
addition, a smartphone version of such a system is currently under Morgenstern 1944), and it is designed to include the contributions
development (Faulkner et al. 2011). of lead time and uncertainties from the selected models and EEW
Japans EEW system, hosted by the Japan meteorological agent information in the automated decision making. ePAD provides a
(JMA), has been broadcasting warnings since 2007 (Allen et al. flexible platform for application-specific models in the decision-
2009b). The warning can be received through cellphones, televi- making process, and allows users to select structural and decision
sions, and the Internet. Particularly for applications to tall build- models that match their desired decision behavior.
ings, Kubo et al. (2011) investigated a method that combines an ePAD makes decisions based on a cost-benefit approach. Apply-
EEW system and a real-time strong motion monitoring system ing a performance-based earthquake early warning (PBEEW)
Fig. 2. Mean of floor acceleration demand: z is the height of the level considered; H is the total height of a building; and α0 is the lateral stiffness ratio
(data from Taghavi-Ardakan 2006)
=1 0
= 3.125 = 12.5
0 0
1 1 1
Fig. 3. Standard deviation of floor acceleration demand: z is the height of the level considered; H is the total height of a building; and α0 is lateral
stiffness ratio (data from Taghavi-Ardakan 2006)
and DFðIM; T lead ; ai Þ ¼ β i ðT lead ÞE½LB jIM; Ai γ 1 as a function of T lead . [β 0 and γ 0 are not needed be-
cause DFðIM; T lead ; a0 Þ ¼ 0].
− γ i ðT lead ÞE½LC jIM; Ai ð7Þ
Similar to the basic model case, DFðIM; T lead ; a0 Þ ¼ 0, Value of Information Model
E½LB jIM; A1 , and E½LC jIM; A1 are the same as in the basic If an EEW system continually provides early warning information,
model, and the discounting factors β and γ as a function of T lead one may want to delay a decision of initiating a mitigation action
are chosen as follows:
when the expected warning lead time is sufficiently long and the
Let T a be the amount of time required to finish the action. When
T lead < T a , the elevator cannot appropriately stop at the closest
floor and open the door for people to evacuate. Thus, β 1 is assumed
to be a step function with value of zero when T lead < T a , and with
value of 1 otherwise. This represents an all-or-nothing benefit
model. For the time delay and service interruption cost, γ 1 is
assumed to be a simple linear function with value between r0
and 1 when T lead < T a , where r0 represents the ratio of fixed
cost (independent of T lead ) over the total cost. In this study, the
service interruption cost rC2 lC2 is taken as a fixed cost, thus,
Fig. 4. β 1 and γ 1 as a function of T lead
r0 ¼ rC2 lC2 =ðrC1 lC1 þ rC2 lC2 Þ. Fig. 4 shows the value of β 1 and
Δt
LetpðIMjDðtÞÞ ¼ ϕ ð9Þ
− Δt; a1 Þ; 0p½IMjDðtÞp½T lead jDðtÞdIMdT lead ð8Þ σIM
Here, Δt is the expected time interval for the next EEW update. and PðDMjIM; A1 Þ ¼ Pðln PFA > ln pfa0 j ln PGAÞ ð10Þ
With the max function involved, VoI often does not have a closed
form solution and so a numerical approximation must be made.
ln PGA þ μST − ln pfa0
This approximation can be prelearned by a surrogate model so that ¼Φ
it can be rapidly computed as the EEW information arrives. In this σST
Z
study, the RVM from machine learning (Tipping 2001) is chosen to
then P½DMjDðtÞ; A1 ¼ PðDMjIM; A1 Þp½IMjDðtÞdIM
construct a surrogate model for VoI because of its ability to create a
sparse model using a Bayesian approach. Results are shown in the
μ þ μIM − ln pfa0
section of numerical solution for value of information model using ¼ Φ ST pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi ð11Þ
surrogate model. σ2ST þ σ2IM
When P0 < 0.5, the critical mean value of lnPGA that separates
μ
-0.8 -0.45
-0.85
-0.5
-0.9
-0.55
-0.95
IM
-0.6
σST =0
μ
-1
σST =0.1
Downloaded from ascelibrary.org by LULEA UNIVERSITY OF TECHNOLOGY on 01/20/16. Copyright ASCE. For personal use only; all rights reserved.
-0.65
-1.05
σST =0.2
-0.7
-1.1 σST =0.3
-0.75 σST =0.4
-1.15
σST =0.5
-1.2 -0.8
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
σ IM σ IM
Fig. 6. Decision contours with fixed P0 , μST values and varying σST values (region above the contour represents taking action and below the contour
represents no action)
σST approaches zero, and the contour approaches a horizontal line E½DFjDðtÞ; a1 > 0
when σST dominates σIM . It demonstrates that the decision is less
μST þ μIM − ln pfa0 fγ
sensitive to the IM uncertainty when σST dominates. ⇔Φ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi > þ 1 − r 0 P0 ≜ r T P0
σ2ST þ σ2IM fβ
ð19Þ
Closed Form Solution for Incomplete Action Model
In the incomplete action model, the influence of insufficient lead As before, the decision contour can then be found in closed form
time on decision making is investigated. This model utilizes the by solving Eq. (19) in the case of equality for μIM as a function
lead time information, p½T lead jDðtÞ, from EEW as well. Similar of σIM
to the previous section, Eq. (1) is simplified to
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Take action if and only if E½DFjDðtÞ; a1 > 0 ð14Þ μIM ¼ ðln pfa0 − μST Þ þ 2ðσ2ST þ σ2IM Þerf −1 ð2rT P0 − 1Þ
ð20Þ
Exploiting the separation of T lead and IM in DF for the incom-
plete action model [Eqs. (6) and (7)]
If one rearranges the expression for rT in Eq. (19), it can be
E½DFjDðtÞ; a1 ¼ E½β 1 jDðtÞE½LB jDðtÞ; A1 found that rT ≥ 1, which means that the incomplete action model
in this case introduces an amplification factor rT for P0 from the
− E½γ 1 jDðtÞE½LC jDðtÞ; A1 ð15Þ basic model. In other words, it always gives a more uncertainty-
taking decision. This is because an incomplete action is less
The expected values of β 1 and γ 1 are found by integrating β 1 beneficial than taking the complete action, which leads to a higher
and γ 1 with p½T lead jDðtÞ, which is taken as log-normally distrib- critical μIM value compared to the basic model. When rT P0 ≥ 1,
uted, with mean ln μT and standard deviation σT . [β 1 ðT lead Þ and the action is never taken. In practice, when the basic model sug-
γ 1 ðT lead Þ are shown in Fig. 4] gests taking action, but the incomplete action model suggests the
opposite, then a backup action is triggered instead. Such backup
1 ln T lead − ln μT
Letp½T lead jDðtÞ ¼ ϕ ð16Þ action should be fast to complete, but will usually come with less
T lead σT benefit. In this case, the backup action could be an immediate stop
of the elevator at the current location.
ln μT − ln T a Fig. 7 shows the value of rT as a function of μT and σT . Assum-
then E½β 1 jDðtÞ ¼ Φ ≜ fβ ð17Þ ing T a ¼ 2s and r0 ¼ 0.5 in Fig. 4, rT increases exponentially as
σT
μT approaches and exceeds T a , and its rate of increase reduces as
the uncertainty in the lead time σT increases. On the other hand, the
1 − r0 lnμT þσ2 =2 ln T a − ln μT incomplete action model has no influence on the decision when
and E½γ 1 jDðtÞ ¼ e T Φ − σ T þ r0
Ta σT there is sufficient lead time, where rT ≈ 1.
Fig. 8 shows the decision contours for ln pfa0 ¼ lnð0.5Þ,
ln μT − ln T a
þ ð1 − r0 ÞΦ α0 ¼ 3.125, T 1 ¼ 2.5s, and P0 ¼ 0.3, which represents a typical
σT
dual system high-rise building with around 20–30 floors, and a
≜ fγ þ ð1 − r0 Þf β ð18Þ slightly conservative decision behavior under the U.S. elevator
standard. In this case, μST ¼ 0.82 and σST ¼ 0.22. The parameter
Following a similar approach to that used in the previous section values for lead time models are T a ¼ 2s, r0 ¼ 0.5, σT ¼ 0.2, and
with P0 defined as in Eq. (12), Eq. (14) can be rearranged to be μT ranges from 1 to 4 s. Note that, when μT is less than 2 s,
5
The first integral in Eq. (21) is simply the lognormal
CDF, which can be rewritten as fΦ½ðlnðμT Þ lnðT a þ ΔtÞ=σT g;
rT
ln μT − lnðT a þ ΔtÞ ln μT − ln T a
no-decision contours are shown in the figure. This is because the where rVoI ðμT ; σT Þ ¼ Φ Φ
σT σT
action is never taken due to the lack of benefit of an incomplete
action, which the backup action is taken instead when necessary. ð24Þ
One can observe that as μT decreases, the decision contour deviates Z
from the one without lead time model (same as μT ≈ 3–4 s) and
and I VoI ðμT ; σT Þ ¼ maxf0; PðDMjIM; A1 Þ − P0 g
becomes more uncertainty taking. Consistent with the intuition, this
is due to the reduced expected benefit of an incomplete action.
IM − μIM
×ϕ dIM ð25Þ
σIM
Numerical Solution for Value of Information Model
Using Surrogate Model Note that rVoI I VoI ≥ 0, which means that the value of informa-
tion model in this case increases the value of rT P0 from the incom-
In the value of information model, the influence of adding extra plete action model. In other words, it always gives an even more
benefit for action a0 on decision making is investigated. Before uncertainty-taking decision. This is because the extra benefit from
directly applying a surrogate model to VoI, the authors first sim- delaying the decision represented by VoI leads to an even higher
plify the expression given for the chosen β 1 and γ 1 models in critical μIM value compared to the incomplete action model with no
the section of incomplete action model. Since DFðIM; T lead − waiting for more EEW information. Different from this incomplete
Δt; a1 Þ < 0 when T lead < T a þ Δt, maxfDFðIM; T lead − action model, VoI is not only a function of μT and σT , but is also a
Δt; a1 Þ; 0g ¼ 0 in this range. Hence, the T lead lower bound of function of μIM and σIM . This is reasonable because, for example,
the integral of VoI can be changed to T a þ Δt instead of Δt. Also, one is more likely to delay a decision when the EEW information is
in this range, both β 1 and γ 1 equal 1, which leads to a separation of
the double integral on T lead and IM into two single integrals
T1 = 2.500000 , α 0 = 3.125000, P0CB = 0.300000
-1.45 0.7
μ T =2 0.6
-1.5
μ T =2.5
VOI
0.5
μ T =3
I
-1.55
ln(μ IM)
-1.7
0.2
0.5
-1.75 0.4 0
0.3 -0.5
-1.8 0.2 -1
σ IM
-1.85 0.1 -1.5 μ IM
0 0.1 0.2 0.3 0.4 0.5
σ IM
Fig. 9. RVM surrogate model for I VoI as a function of μIM and σIM
Fig. 8. Decision contours with incomplete action model (T 1 ¼ 2.5, P0 ¼ 0.3 fixed)
Fig. 11. Location of a segment of San Andreas Fault (solid black line) and a chosen building in downtown Los Angeles (star) (map data © 2013
Google, INEGI)
Conclusion
Example
A segment of the San Andreas Fault (from Palmdale to San This study presents a decision framework for elevator control
Bernardino) is selected to demonstrate the authors’ proposed appli- based on uncertain EEW information using the ePAD framework
Downloaded from ascelibrary.org by LULEA UNIVERSITY OF TECHNOLOGY on 01/20/16. Copyright ASCE. For personal use only; all rights reserved.
cation to elevator control. A building with parameters T 1 ¼ 2.5s (Wu et al. 2013) with a suggested structural model, decision model,
and α0 ¼ 3.125 is chosen (same values as previously used in and lead time model. The authors’ analysis shows that instead of
the analyses), that is located in downtown Los Angeles. The solid going through the complication of a complete cost-benefit analysis,
black line in Fig. 11 shows the causative fault segment and the star a user may select the value of a related variable P0 that can re-
indicates the building location. present the users desired type of decision behavior. Also, the influ-
Instead of parameterizing the PDF of IM from EEW, ence of lead time and EEW uncertainties on decisions can be easily
p½IMjDðtÞ, into μIM and σIM , the PDF is found as a function understood through the use of decision contours, which represent
of earthquake magnitude and hypocenter-to-site distance using different types of decision behaviors. A surrogate model using the
the ground motion prediction equation developed by Boore and machine learning algorithm RVM is created for the sake of fast
Atkinson (2008). Hence, instead of plotting the decision contours computation of the value of information model. The steps to setup
in a 2D space of μIM and σIM , the contours are plotted in a 2D space the algorithm for automated decisions of elevator control using un-
of earthquake magnitude and location along the chosen segment of certain EEW information are summarized by:
the San Andreas Fault. Once again, the region above the decision 1. Model parameters setup: Choose T 1 and α0 based on a target
contour represents taking action; the region below the decision con- building; choose Δt and ln pfa0 based on the EEW system
tour represents not to take action. Change of decision contours is operation and the elevator standard for the region where the
investigated based on two factors: varying μT for a fixed ln pfa0 ; building is located; choose T a and r0 based on a specific ele-
and varying ln pfa0 for a fixed μT . vator application;
Fig. 12 shows the result of this simulation. As one would expect, 2. Evaluate decision contour: Pick a P0 value and generate de-
the location on the fault line that is the closest to the building has cision contours as a function of (μT , σT ) and (μIM , σIM ) or
the lowest critical magnitude to trigger the action, and a decrease in (earthquake magnitude, fault line segments); and
ln pfa0 lowers the whole-decision contour in this plot. Note that 3. Check decision behavior: Iteratively increase or decrease P0 to
no-decision contour is shown in the plot for ln pfa0 ¼ lnð0.5Þ find a value that can best represent the user’s desired decision
(the U.S. standard) because the action is never taken within the behavior under uncertain EEW information.
tested range of magnitude on the chosen fault segment. This is Further refinement of the models employed in ePAD for elevator
not an intuitive result and the authors suggest that the current stan- control could be explored. As explained in Wu et al. (2013), more
dard may need to be changed if a robust decision framework is to be flexibility in the types of decision behavior can be obtained by ad-
put in place. Also, μT ¼ 2.5s leads to the lowest set of values of the justing Δt. Although the chosen structural model can be readily
critical magnitude along the fault segment because after including applied to various types of buildings, in order to obtain a better
the lead time uncertainty, this is the mean lead time that has the least representation of a target building, a more sophisticated structural
model can be used to replace the simpler model used in this paper.
If necessary, the decision model and lead time model (β 1 and γ 1 )
9 can also be changed, at the possible expense of introducing extra
Jap 0.15g μ T =2 complexity into the problem.
Jap 0.15g μ T = 2.5
8.5
Jap 0.15g μ =3
T
Jap 0.15g μ T = 3.5
8 Jap 0.15g μ T =4 Acknowledgments
Earthquake Magnitude
6.5
References
6
Allen, R., Brown, H., Hellweg, M., Khainovski, O., Lombard, P., and
Neuhauser, D. (2009a). “Real-time earthquake detection and hazard
5.5
assessment by ElarmS across California.” Geophys. Res. Lett., 36,
L00B08.
5 Allen, R., Gasparini, P., Kamigaichi, O., and Böse, M. (2009b). “The status
Palmdale to San Bernardino
of earthquake early warning around the world: An introductory over-
Fig. 12. Decision contour of a dual system building (T 1 ¼ 2.5s) view.” Seismol. Res. Lett., 80(5), 682–693.
in downtown Los Angeles for earthquake magnitude between Boore, D., and Atkinson, G. (2008). “Ground-motion prediction equations
for the average horizontal component of PGA, PGV, and 5%-damped
M5–M8 on a segment of San Andreas Fault between Palmdale and
PSA at spectral periods between 0.01 s and 10.0 s.” Earthquake
San Bernardino
Spectra, 24(1), 99–138.
Soc. Am., 99(2A), 897–905. of Steel Structures in Seismic Area (STESSA), CRC Press, Boca Raton,
Cal/OSHA (Division of Occupational Safety and Health). (2012). “Title 8 FL.
regulations, sub-chapter 6, elevator safety orders, group 3, new elevator Kubo, T., Hisada, Y., Murakami, M., Kosuge, F., and Hamano, K. (2011).
installations, article 37, seismic requirements for elevators, escalators “Applications of an earthquake early warning system and a real-time
and moving walks, section 3137.” 〈http://www.dir.ca.gov/Title8/3137 strong motion monitoring system in emergency response in a high-rise
.html〉 (Apr. 27, 2014). building.” Soil Dyn. Earthquake Eng., 31(2), 231–239.
Campbell, K., and Bozorgnia, Y. (2008). “NGA ground motion model for Miranda, E., and Reyes, C. (2002). “Approximate lateral drift demands
the geometric mean horizontal component of PGA, PGV, PGD and 5% in multi-story buildings with nonuniform stiffness.” J. Struct. Eng.,
damped linear elastic response spectra for periods ranging from 0.01 to 10.1061/(ASCE)0733-9445(2002)128:7(840), 840–849.
10 s.” Earthquake Spectra, 24(1), 139–171. Miranda, E., and Taghavi-Ardakan, S. (2005). “Approximate floor accel-
Cheng, M., Wu, S., Heaton, T., and Beck, J. (2014). “Earthquake early eration demands in multistory buildings. I. Forumulation.” J. Struct.
warning application to buildings.” Eng. Struct., 60, 155–164. Eng., 10.1061/(ASCE)0733-9445(2005)131:2(203), 203–211.
Cua, G., and Heaton, T. (2007). “The virtual seismologist (VS) method: A Raiffa, H., and Schlaifer, R. (1961). Applied statistical decision theory,
Bayesian approach to earthquake early warning.” Earthquake early Massachusetts Institute of Technology Press, Cambridge, MA.
warning systems, P. Gasparini, G. Manfredi, and J. Zschau, eds., Taghavi-Ardakan, S. (2006). “Probabilistic seismic assessment of floor ac-
Springer, Naples, Italy, 97–132. celeration demands in multi-story buildings.” Ph.D. thesis, Stanford
Faulkner, M., Olson, M., Chandy, R., Krause, J., Chandy, K., and Krause, Univ., Stanford, CA.
A. (2011). “The next big one: Detecting earthquakes and other rare Tipping, M. (2001). “Sparse Bayesian learning and the relevance vector
events from community-based sensors.” Proc., 10th ACM/IEEE Int. machine.” J. Mach. Learn. Res., 1(3), 211–244.
Conf. on Information Processing in Sensor Networks, IEEE, New York, Von Neumann, J., and Morgenstern, O. (1944). Theory of games and
13–24. economic behavior, Princeton University Press, Princeton, NJ.
Grasso, V. (2005). “Seismic early warning systems: Procedure for auto- Wu, S., Beck, J., and Heaton, T. (2013). “ePAD: Earthquake probability-
mated decision making.” Ph.D. thesis, Università degli Studi di Napoli based automated decision-making framework for earthquake early
Federico II, Napoli. warning.” Comput. -Aided Civ. Infrastruct. Eng., 28(10), 737–752.
Hilbring, D., Titzschkau, T., Buchmann, A., Bonn, G., Wenzel, F., and Yahoo News. (2013). “Strong earthquake rattles Taiwan, no reports of
Hohnecker, E. (2014). “Earthquake early warning for transport lines.” casualties.” 〈http://news.yahoo.com/earthquake-6-7-magnitude-strikes-
Nat. Hazards, 70(3), 1795–1825. taiwan-usgs-122240396.html〉 (Apr. 9, 2014).