Reliability Engineering and System Safety 99 (2012) 24–32

Contents lists available at SciVerse ScienceDirect

Reliability Engineering and System Safety

journal homepage: www.elsevier.com/locate/ress

Study on probability distribution of fire scenarios in risk assessment to

emergency evacuation
Chu Guanquan a,n, Wang Jinhui b
a
Waterborne Transportation Institute, Ministry of Transport, 8 Xi Tu Cheng Road, Haidian District, Beijing 100088, China
b
College of Ocean Environment and Engineering, Shanghai Maritime University, Shanghai 201306, China

a r t i c l e i n f o a b s t r a c t

Article history: Event tree analysis (ETA) is a frequently-used technique to analyze the probability of probable fire
Received 11 February 2011 scenario. The event probability is usually characterized by definite value. It is not appropriate to use
Received in revised form definite value as these estimates may be the result of poor quality statistics and limited knowledge.
24 October 2011
Without addressing uncertainties, ETA will give imprecise results. The credibility of risk assessment
Accepted 30 October 2011
will be undermined. This paper presents an approach to address event probability uncertainties and
Available online 6 November 2011
analyze probability distribution of probable fire scenario. ETA is performed to construct probable fire
Keywords: scenarios. The activation time of every event is characterized as stochastic variable by considering
Risk assessment uncertainties of fire growth rate and other input variables. To obtain probability distribution of
Fire scenario
probable fire scenario, Markov Chain is proposed to combine with ETA. To demonstrate the approach, a
Event tree
case study is presented.
Monte Carlo simulation
Evacuation & 2011 Elsevier Ltd. All rights reserved.

1. Introduction protection design in essence is to reduce the risk of life and/or

Fire has caused a large number of people’s death in many
public assembly occupancies such as commercial buildings, hotels
and public amusement places in recent years. For example of
China, fatalities happened occasionally in these buildings after
entering the 21st century. Fire occurred at Dongdu Mansion in
Luoyang, Henan province caused 309 fatalities on 25 Dec, 2000.
Fire occurred at Zhongbai Mansion in Jilin, Jilin province caused
56 death and 70 injuries on 15 Feb, 2004. Fire occurred at a club in
Shenzhen, Guangdong province caused 44 death and 88 injuries
on 20 Sep, 2008. To ensure occupant life safety, fire safety design
and management are two indispensable measures. The former is
generally concerned with the design phase of buildings. The latter
is mainly concerned with the occupancy phase of buildings. Since
most of these public assembly occupancies are large scale build-
ings, high-rise buildings or underground buildings, performance-
based fire protection design has become increasingly accepted as
an alternative solution to the prescriptive-based design [1,2]. In
the current performance-based fire protection design, the concept
of risk plays a particularly important role in quantifying fire safety
levels [3]. Risk assessment approaches have become gradually
extensive application in fire safety design for solution of practical
engineering problems. The objective of performance-based fire
n
Corresponding author. Tel./fax: þ 861062364733.
E-mail address: chugq@wti.ac.cn (C. Guanquan).
property. An appropriate performance-based fire protection
design cannot be well implemented without reasonable fire risk
assessment. Moreover, to enhance fire protection management
level, fire public liability insurance is a wonderful approach. Fire
risk assessment in terms of expected risk to life is the basis for
determining fire public liability insurance rate. The credible fire
public liability insurance rate requires reasonable risk assessment
results. Then, not only performance-based fire protection design,
but fire public liability insurance rate determination requires
quantitative fire risk assessment.
It has been proposed that fire risk is an undesirable conse-
quence potentially induced by a fire [4]. Fire risk is usually viewed
as a numerical value that is a function of probability and
consequence [5]. The scenario-based approach is usually adopted
in risk assessment process [6–8]. In scenario-based fire risk
assessment, there are two main evaluation parameters, namely,
probability and corresponding consequence. The current studies
usually focus on evaluating consequence, namely number of
deaths and injuries based on ASET/RSET timeline assessment
schematically shown in Fig. 1. If RSET (Required Safe Egress Time)
is larger than ASET (Available Safe Egress Time), some people
cannot evacuate to safety zone before the onset of untenable
conditions. The fatalities may occur at ‘‘Risk Interval’’.
When evaluating consequence of a fire scenario, uncertainties
of fire dynamics and occupant evacuation time are usually
considered in calculating ASET and RSET [9–12]. In these refer-
ences, uncertainties of consequences are extensively addressed.

0951-8320/$ - see front matter & 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ress.2011.10.014
 C. Guanquan, W. Jinhui / Reliability Engineering and System Safety 99 (2012) 24–32
However, when evaluating probability of probable fire scenario,
little attention is paid to uncertainties. To obtain reasonable
results, uncertainty of probability of probable fire scenario should
also be considered in the risk assessment process. When evaluat-
ing probabilities of fire scenarios, various fire scenarios are
usually constructed by an event tree by considering fire protec-
tion system and manual intervention as events. Traditionally,
probability of every event is usually estimated to calculate the
outcome branch probability. In practice, it is not appropriate to
make a precise estimation of event probability because these
estimates may be the result of poor quality statistical data and
experts’ limited knowledge.
In fact, the probability of a fire protection system is dominated
by two factors, namely operational reliability and activation
probability. Operational reliability is the basic premise that a fire
protection system can operate successfully. Moreover, the activa-
tion probability also plays an important role besides operational
reliability. Fire will be detected and controlled more easily if a fire
protection system activates earlier. Generally, the activation time
of a fire protection system is dominated by fire dynamics and
building environment. Since the type of combustible materials,
their arrangement and source of ignition is unpredictable, the
activation time of a fire protection system is also a random
variable. That is to say, activation probability is a variable with
time. Then, event tree will propagate the uncertainty of event
probability for the output, namely probability of probable fire
scenario. If probability of a fire scenario can be achieved as a
Fig. 1. ASET/RSET timeline.
Automatic Manual
Fire Manual
fire fire
ignition extinguishment
detection detection
Yes
Yes
No
Yes
Yes
No
No
No
Fig. 2. Event tree analysis on probable fire scenarios.
25
variable with time elapsing, the corresponding probability at Risk
Interval will be easily obtained.
The main focus of this paper is to describe a stochastic
approach refer to evaluating probability distribution of fire
scenarios at different time. The core of the approach is that
stochastic activation times of fire protection system and manual
intervention are used to estimate probability distribution of
probable fire scenario. To analyze variation of the event prob-
ability with time, the Monte Carlo method is coupled with fire
models. To describe the process that the event probabilities being
propagated into the different branches of the tree, event tree is
coupled with the Markov Chain. The following sections describe
the steps to analyze variation of probability of probable fire
scenario with time. In Section 2 of the paper, probable fire
scenarios are analyzed by constructing an event tree. Uncertainty
of design fires is proposed by considering lognormal fire growth
rates. The activation time of fire protection system or manual
intervention is analyzed by a coupling fire model and the Monte
Carlo simulation. In Section 3 of the paper, probability distribu-
tion of probable fire scenario is analyzed using the Markov Chain
and event tree. In Section 4 of the paper, a case study is proposed
to explain the approach in detail. The last section of the article
presents the conclusion and discussion in the future study.
2. Event tree analysis on probable fire scenario and treatment
of uncertainty
In fire risk assessment, it may be necessary to structure a large
number of possible fire scenarios arising from an initial event.
Generally, evaluating probability of a fire scenario involves deal-
ing with two various uncertainties, namely, stochastic uncertainty
and epistemic uncertainty [13]. Stochastic uncertainty is the
uncertainty that regards the outcome of a fire. Epistemic uncer-
tainty is the uncertainty of the values of the variables used in the
Automatic
Scenario State Outcome
Sprinkler
1 7 No Fire
Yes
2 6 No Fire
No
3 5 Fire
4 4 No Fire
Yes
5 3 No Fire
No
6 2 Fire
Yes
7 1 No Fire
No
8 0 Fire
26 C. Guanquan, W. Jinhui / Reliability Engineering and System Safety 99 (2012) 24–32
model for simulating a given fire scenario. The methods of dealing
with the two kinds of uncertainties are described in the following
subsections.
2.1. Probable fire scenarios constructed by event tree
Event tree analysis is usually utilized to deal with stochastic
uncertainty by constructing probable fire scenarios in fire risk
assessment. The event tree for the current study is illustrated in
Fig. 2. Four possible events, namely, automatic fire detection,
manual fire detection, manual extinguishment and automatic
sprinkler activation, are considered in the event tree. If prob-
ability of every event is a variable with time, they will be
propagated into the different branches of the tree. Then, prob-
ability of probable fire scenario at different time can be easily
obtained by the approach proposed in this study.
2.2. Treating uncertainty of design fires
To calculate the probability of every event at different time,
epistemic uncertainty should be considered. A design fire is
generally considered to describe the main time-varying proper-
ties of a fire based on reasonable assumptions about the type and
quantity of combustibles, ignition source, growth and decay of the
fire [14]. The main time-varying property that defines a fire is the
heat release rate. The heat release rate of a fire at initial phase is
often approximated by a time squared growth form.
Q ¼ at 2
where Q is heat release rate (kW); a is fire growth rate (kW/s );
t is the time after fire ignition (s).
Fire growth rate is often categorized discretely into low,
medium, fast and ultra-fast levels with the corresponding values
of 0.00293, 0.01172, 0.0469 and 0.1876 kW/s2, respectively [15].
The corresponding heat release rate curves are illustrated graphi-
cally in Fig. 3. Due to the extreme variability of fire in buildings,
the limited discrete fire growth rates may not be sufficient to
characterize all the possible fire scenarios for risk-based assess-
ment. In reality, fire growth rate may vary in a wide range
depending on some random parameters, such as the fuel type
and configuration, the point of ignition and the ventilation
condition, etc. Therefore, fire growth rate may exhibit itself as a
continuous random variable. Based on collection from real fire
95th percentile Ultra-fast Fast
Medium Slow
5000
4000
Mean Value
Heat Release Rate/kW
3000
50th percentile
(Median Value)
2000
1000
0 pðti1,i 9t 0,i1 Þ ¼
0 200 400 600
Time from ignition/s
Fig. 3. Heat Release Rate curves of t2 fire with time from ignition.
incidents, the observed fire growth rates were reasonably well
approximated by the lognormal distribution [16]. For instance,
the mean value, the median and 95th percentile of a lognormal
fire growth rate in retail occupancy building is 0.027, 0.0045 and
0.103 kW/s2, respectively. These three values are illustrated in
Fig. 3. In this study, a range of design fires is proposed with
consideration of lognormal fire growth rates.
2.3. Treating uncertainty of event probability
It is important to consider uncertainty of event probability as
event tree analysis will propagate them to outcome events. In
general, the probability of a fire protection system is mainly
dominated by two factors, namely operational reliability and activa-
tion probability. For example, smoke detector and sprinkler will
operate as long as the reliability can be ensured. The activation time
of fire detection and sprinkler can be considered as a cumulative
probability distribution. Moreover, when fire detector fails to
activate, manual detection may influence fire development. Even if
it is difficult for the occupants to be aware of fire cues in the fire
initial phase, occupants in the building will detect fire cues with
time elapsing. The manual detection time can also be considered as
a cumulative probability distribution. The cumulative probability
distribution cannot be directly used as branching probability in the
event tree analysis. Suppose a smoke detector is activated at time
interval [ti 1,ti], it is equivalent that fire is not detected before time
ti 1. Probability of every event in the event tree can be obtained by
ð1Þ Pde ¼ Rde pde ðt i1,i 9t 0,i1 Þ ð2Þ
2
Psp ¼ Rsp psp ðt i1,i 9t 0,i1 Þ ð3Þ
Pmd ¼ pmd ðt i1,i 9t 0,i1 Þ ð4Þ
where Pde is probability of automatic fire detection; Psp is probability
of automatic sprinkler; Pmd is probability of manual fire detection.
Rde is the operational reliability of automatic fire detection; Rsp is the
operational reliability of automatic sprinkler; ti 1,i is the corre-
sponding event that is being activated at time interval [ti 1,ti]; t 0,i1
is the corresponding event that is not being activated at time
interval [t0,ti 1]. pde ðt i1,i 9t 0,i1 Þ is the conditional probability of
automatic fire detection; psp ðt i1,i 9t 0,i1 Þ is the conditional prob-
ability of automatic sprinkler; pmd ðt i1,i 9t 0,i1 Þ is the conditional
probability of manual fire detection.
In general, operational reliability can be obtained based on fire
statistics. The remaining task is to achieve conditional probability
of every event. Suppose there are n time intervals, namely
[t0,t1],y,[ti  1,ti],y,[tn  1,tn] in the considered time-dependent
fire process. Based on the Bayes’ theorem
pðt i1,i 9t 0,i1 Þ
pðt i1,i Þpðt 0,i1 9t i1,i Þ
¼
pðt 0,1 Þpðt 0,i1 9t 0,1 Þ þ    þpðt i1,i Þpðt 0,i1 9t i1,i Þ þ    þ pðt n1,n Þpðt 0,i1 9t n1,n Þ
ð5Þ
where pðt i1,i 9t 0,i1 Þ is the conditional probability that the corre-
sponding event occurs at time interval [ti  1,ti].
For a random time tj, an expression of the denominator in Eq. (5) is
(
0 joi
pðt 0,i1 9t j1,j Þ ¼ : ð6Þ
1 jZi
Substituting Eq. (6) in Eq. (5)
pðt i1,i Þ Fðt i ÞFðt i1 Þ
¼ ð7Þ
pðt i1,i Þ þ    þ pðt n1,n Þ 1Fðt i1 Þ
800
where F(ti) is the cumulative probability function of activation
time or manual detection time at time ti. It can be achieved by
coupling fire models and the Monte Carlo simulation.
 C. Guanquan, W. Jinhui / Reliability Engineering and System Safety 99 (2012) 24–32 27

After finding fire cues, occupants may try to use fire extin- P11
guisher to control and extinguish fires. In general, whether
fire can be extinguished manually is mainly dominated by the
Heat Release Rate (HRR). The critical HRR of fires can be
extinguished manually is 950 kW [17]. Probability of manual 1
extinguishment is
P06 P03
N HRR r 950 ðt i Þ P66 6 3 P33
Pe ¼ ð8Þ P01
N total

where NHRR r 950 is the number of fire curves, which heat release P56 P23
rate is smaller than 950 kW at time ti; Ntotal is the number of
design fire curves.
P05 P02
P55 5 0 2 P22

2.4. Coupling Monte Carlo simulation with fire models

P57 P24
If stochastic activation time is available, cumulative probabil- P00
ity distribution of every event can be obtained according to
P07 P04
statistical theory. In this study, the Monte Carlo simulation is P77 P44
7 4
used to address uncertainty of inputs to stochastic outputs. A
program has been developed, which combines the Monte Carlo
simulation technique with a two-zone fire model for single
Fig. 4. Transition diagram of every state.
compartments. In the Monte Carlo method, a large number of
samples of input parameters are randomly chosen according to
the corresponding probability distributions. After thousands of
fire scenario 8 to 1. Based on transition of every state, transition
iterations, the calculation result can be obtained, such as activa-
diagram is shown in Fig. 4.
tion times of fire protection system and manual intervention. The
Suppose Pi,j is the transition probability from state i to j. When
result is subjected to statistical analysis to obtain probability
Pi,j ¼0, it means that there is no edge connecting state i with j.
distribution of activation time. Moreover, sensitivity of the output
When Pi,j 40, there is a transition probability from state i to j.
variables to the input variables is also analyzed in terms of the
When Pi,i ¼1, it means that there is no leaving node i to other
rank order correlations.
nodes besides self-loop.
In this study, fire detection time is calculated based on the
Based on the transition diagram shown in Fig. 4, transition
criterion that detectors are activated as smoke layer being below
the ceiling by 5% of the ceiling height [6]. The sprinkler activation matrix P of the transition diagram is
0 1
time is calculated by DETACT-QS [18]. The manual detection time p0,0 p0,1    p0,7
is assumed to occur when smoke layer descends below the ceiling Bp C
B 1,0 p1,1    p1,7 C
by 10% of the ceiling height. Besides lognormal fire growth rate, P¼B B ^
C: ð9Þ
@ ^ ^ ^ C
A
uncertainties of other variables in the sprinkler activation model p7,0 p7,1    p7,7
are also considered, such as sprinkler Response Time Index (RTI),
sprinkler activation temperature, ambient temperature and dis- Substituting the probability of every event in Eq. (9), transition
tance between fire origin and detector, etc. matrix can be expressed as

0 1
P d Pmd Psp P d P md P sp P d Pmd Pe P sp Pd Pmd Pe P sp Pd P md P e P d Pe P sp P d P e Psp Pd Pe
B C
B 0 1 0 0 0 0 0 0 C
B C
B 0 0 Pe P sp Pe P sp Pe 0 0 0 C
B C
B C
B 0 0 0 1 0 0 0 0 C
P¼B
B
C: ð10Þ
B 0 0 0 0 1 0 0 0 CC
B C
B 0 0 0 0 0 Pe P sp Pe P sp Pe C
B C
B 0 0 0 0 0 0 1 0 C
@ A
0 0 0 0 0 0 0 1

3. Variation of probability of probable fire scenario with time

To analyze variation of probability of probable fire scenario
with time, the Markov Chain is proposed to combine with event
tree. The time evolution of a fire incident determined from event
tree can be treated as a Markov-process [19]. Scenario 8 of the
event tree shown in Fig. 2 can be considered as the initial state
because fire protection systems and manual intervention are not
successful, which is equivalent to no fire breaking out at time
zero. The eight states are assigned from 0 to 7 corresponding with
Suppose the probability vector at a random time ti is
PðSti Þ ¼ ðPðsti ,0 Þ,Pðsti ,1 Þ,. . ., Pðsti ,7 ÞÞ: ð11Þ
Probability vector at next time ti þ 1 can be expressed as
PðSti þ 1 Þ ¼ PðSti Þ  Pti þ 1 ð12Þ
where Pti þ 1 is the transition matrix from time ti to time ti þ 1.
Suppose the time-dependent fire process is divided into n
time intervals. Based on Eq. (12), probability vectors at the later
28 C. Guanquan, W. Jinhui / Reliability Engineering and System Safety 99 (2012) 24–32

STEP 1: To analyze probable fire scenario

Automatic fire Manual fire Manual Automatic

detection detection extinguishment Sprinkler

Event tree

Probable fire scenario

STEP 2: Stochastic analysis on probability of every event with time

Fire dynamics model

Uncertainty of design fires Two-zone fire model &
(e.g. DETACT-QS, two-zone
(Lognormal fire growth rate) Evacuation model
fire model)

Monte Carlo simulation

Stochastic analysis on probability

of every event with time

STEP 3: Probability distribution of probable fire scenario

Stochastic analysis on probability

Probable fire scenario
of every event with time

Markov Chain

Time- varying probability of

probable fire scenario

Fig. 5. The schematic framework of analyzing probability distribution of probable fire scenario.

time are according to the National Standard for Fire-Safety Design of

PðSt1 Þ ¼ PðSt0 Þ  Pt1
PðSt2 Þ ¼ PðSt1 Þ  Pt2 ¼ PðSt0 Þ  Pt1  Pt2
^
PðSti Þ ¼ PðSti1 Þ  Pti ¼ PðSt0 Þ  Pt1  Pt2      Pti
^ 4.1. Input variables and relative fire models
PðStn1 Þ ¼ PðStn2 Þ  Ptn1 ¼ PðSt0 Þ  Pt1  Pt2      Ptn1
PðStn Þ ¼ PðStn1 Þ  Ptn ¼ PðSt0 Þ  Pt1  Pt2      Ptn1  Ptn :
Since fire scenario 8 is considered as the initial state in the
system, the probability vector at initial time is P(St0) ¼(1,0,y,0).
Moreover, event probability at different time can be also obtained
by coupling fire models with the Monte Carlo simulation. Transi-
tion matrix from random time ti to time ti þ 1 can also be achieved
by Eq. (10). Then, probability vector, namely probability of
probable fire scenario at different time can be calculated accord-
ing to Eq. (13).
To express the analysis process, a flow diagram describing the
process is shown in Fig. 5. Step 1 is to analyze probable fire
scenario by event tree. Step 2 is to analyze uncertainty of
probability of every event. Step 3 is to obtain time-varying
probability of probable fire scenario by the Markov chain.
4. Case study
A case study is presented to express the approach in detail. The
building environment is a compartment of a commercial building
Buildings, Code of Design on Building Fire Protection and Preven-
tion (GB50016-2006). The fire compartment is 40 m long, 30 m
wide and 3.5 m high. There are two exits with dimensions of 4 m
in width and 2.5 m high from the ground.
ð13Þ In this case study, lognormal fire growth rate is considered to
express uncertainty of design fires. The activation times of fire
protection systems and manual intervention are analyzed by
deterministic models with consideration of random variables.
The activation time estimation models are shown in Table 1. A
complete list of the random variables in the deterministic models
is given in Table 2. From life safety point of view, the most
important time period in a building fire is the first 10 or 15 min
during which the occupants may evacuate to safety [17]. The
time-dependent fire process in this case study is proposed as
600 s.
4.2. Stochastic analysis on activation time of fire protection system
and manual intervention
After 5000 iterations of the Monte Carlo simulations, the
histogram of fire protection system and manual intervention is
shown in Fig. 6. The time interval is chosen as 10 s. The opera-
tional reliability of fire protection system is shown in Table 3.
Based on the Bayes’ theorem and cumulative histogram in Fig. 6,
 C. Guanquan, W. Jinhui / Reliability Engineering and System Safety 99 (2012) 24–32 29

Table 1
The activation probability/time of every event.

Event Calculation model Expression

Automatic fire detection
Manual fire detection
Manual extinguishment
Automatic sprinkler
Fire zone model (smoke layer being below the ceiling by 5% of the ceiling height) Cumulative probability distribution
Fire zone model (smoke layer being below the ceiling by 10% of the ceiling height) Cumulative probability distribution
N HRR r 950 ðti Þ The critical HRR is 950 kW
Pe ¼ N total
DETACT-QS Cumulative probability distribution

Table 2
A list of random variables used in fire models.

Variable Distribution Mean Standard deviation Min Max Unit

Fire growth rate Lognormal 0.0275 0.165 / / kW/s2

Response Time Index (RTI) Normal 90 10 / / (m/s)1/2
Sprinkler activation temperature Normal 68 5 / / 1C
Ambient temperature Normal 24 2 / / 1C
Distance between fire origin and detector Uniform / / 0 2.83 m

100% 100%

75% 75%
Percent

Percent

50% 50%

25% 25%

0% 0%
20 40 60 80 100 120 140 160 180 200 100 125 150 175 200 225 250 275 300
Smoke detection time/s Manual detection time/s

100%
Probability of manual extinguishment

1.0

75% 0.8
Percent

0.6
50%

0.4

25%
0.2

0% 0.0
0 50 100 150 200 250 300 350 400 0 100 200 300 400 500 600
Sprinkler activation time/s time/s

Fig. 6. The histogram of smoke detection time (a), manual detection time (b) sprinkler activation time (c) and manual extinguishment time (d).

Table 3
probability of smoke detection, manual detection, sprinkler and
Statistical operational reliability of fire protection
system. manual extinguishment can be calculated by Eqs. (7) and (8).

Fire protection system Operational reliability 4.3. Probability distribution of probable fire scenario

Sprinkler 0.93
Automatic detection 0.72
Based on event probability at different time, transition matrix
of Eq. (10) is available to calculate the probability vector
30 C. Guanquan, W. Jinhui / Reliability Engineering and System Safety 99 (2012) 24–32

according to Eq. (13). Thus, probability distribution of probable
fire scenario is shown in Fig. 7. The histogram of Fig. 6(a), (b), (d),
(e) and (g) demonstrates that probability of fire scenario 1, 2, 4,
5 and 7 are increasing to a constant with time elapsing. The
histogram of Fig. 7 (c) and (f) demonstrates that probability of fire
scenario 3 and 6 increases to a maximum and decreases with time
elapsing. The histogram of Fig. 7(h) demonstrates that probability
of fire scenario 8 decreases with time elapsing.
The dominated time interval corresponding to probability
value plays an important role in risk assessment for emergency
evacuation as fatalities may occur at the risk interval shown in
Fig. 1. The plots of Fig. 7 show that probability of fire scenario is

70 600 70 600
0.6 Scenario 1 0.010 Scenario 2
0.5
0.008
Probability

Probability
0.4
0.006
0.3
0.004
0.2
0.1 0.002

0.0 0.000
0 100 200 300 400 500 600 0 100 200 300 400 500 600
time/s time/s

70 230 100 600

0.012 0.09
Scenario 3 0.08 Scenario 4
0.010
0.07
0.008 0.06
Probability

Probability

0.05
0.006
0.04
0.004 0.03
0.02
0.002
0.01
0.000 0.00
0 100 200 300 400 500 600 0 100 200 300 400 500 600
time/s time/s

100 600 100 210

0.0016
Scenario 5 0.0020 Scenario 6
0.0014
0.0012
0.0015
Probability

Probability

0.0010
0.0008
0.0010
0.0006
0.0004 0.0005
0.0002
0.0000 0.0000
0 100 200 300 400 500 600 0 100 200 300 400 500 600
time/s time/s

40 600 0 180
0.5 1.0
Scenario 7 Scenario 8
0.4 0.8
Probability

Probability

0.3 0.6

0.2 0.4

0.1 0.2

0.0 0.0
0 100 200 300 400 500 600 0 100 200 300 400 500 600
time/s time/s

Fig. 7. Probability distribution of probable fire scenario.

 C. Guanquan, W. Jinhui / Reliability Engineering and System Safety 99 (2012) 24–32
Manual extingsuishment
Smoke detection
6
Fire scenario
-1.0 -0.8 -0.6 -0.4 -0.2 0.0
Rank correlation coefficient for probability of fire scenarios
Fig. 8. The sensitivity of event probability to probability of probable fire scenario.
dominated by from scenario 1 to scenario 8 in early fire phase.
However, as time increases, the time interval of fire scenario 3,
6 and 8 is smaller than that of other fire scenarios. Probability of
fire scenario is mainly dominated by scenario 1, 2, 4, 5 and
7 irrespective of scenario 3, 6 and 8.
4.4. Sensitivity analysis
To analyze the effect of event probability on likelihood of
probable fire scenario, the sensitivity is presented by calculating
the Spearman’s rank-order correlation coefficients (RCC). The
sensitivity analysis result is shown in Fig. 8. The most important
event in fire scenario 1, 4, 5 and 6 is manual extinguishment. The
most important event in fire scenario 2 and 3 is smoke detection.
The most important event in fire scenario 7 and 8 is sprinkler.
Generally, the plots shown in Fig. 8 demonstrate that the most
influential event managed probability of fire scenario is the
sprinkler.
5. Discussion
When performing fire risk assessment to life safety under
emergency evacuation, it is expected to obtain more precise
results. If ASET oRSET at a fire scenario, casualties may occur.
The ‘‘Risk Interval’’, namely [ASET,RSET] is the concerned focus in
risk assessment to emergency evacuation. Based on the approach
proposed in this article, probability of probable fire scenario can
be evaluated as a time-varying variable, such as the plots shown
in Fig. 7. Then, the corresponding probability at time interval
[ASET,RSET] can be achieved. That is to say, probability of escape
failure under probable fire scenario can be obtained. If the
31
Manual detection
Sprinkler
0.2 0.4 0.6 0.8 1.0
corresponding consequence is available, fire risk to emergency
evacuation will be evaluated.
To analyze probability distribution of probable fire scenario in
emergency evacuation is a complex process, which involves a
large number of uncertainties from fire dynamics, human beha-
viors and building environment. Some key aspects should be
concerned as using this approach. The basic is to construct the
event tree. Although it seems that it is easy to construct an event
tree by considering fire protection and manual intervention, some
event, especially important event cannot be omitted. For example,
sprinkler is one of these. Historical records reveal that fatalities
will hardly occur as the sprinklers are operational. Moreover,
uncertainty of design fires should be considered by reasonable
assumptions and reliable data.
6. Conclusion
This paper presents an approach to analyze probability dis-
tribution of fire scenarios in risk assessment for emergency
evacuation. The approach can give a stochastic analysis on time-
dependent probability of probable fire scenario. When construct-
ing probable fire scenarios, event tree is used with consideration
of fire protection system and manual intervention. The probabil-
ity of every event is not treated as a definite value but a stochastic
parameter. The activation time of every event is characterized as a
stochastic variable by coupling the Monte Carlo approach and fire
models according to uncertainties of fire growth rate and other
input variables. To achieve probability distribution of probable
fire scenario at different times, the Markov Chain is used to
combine with event tree. To express the approach in detail, a case
study is presented. Sensitivity analysis is analyzed in terms of the
rank order correlations. To improve the approach, the further
32 C. Guanquan, W. Jinhui / Reliability Engineering and System Safety 99 (2012) 24–32
work will focus on two aspects. First, the assumption of manual
detection time will be analyzed by more reasonable mathematical
model. Second, the program, which combines the Monte Carlo
method with a two-zone fire model for single compartments will
be developed for multi-enclosure.
Acknowledgment
This paper is sponsored by the National Natural Science
Foundation of China (Grant nos. 70803015 and 50909058) and
the ‘Chen Guang’ project of Shanghai Municipal Education Com-
mission and Shanghai Education Development Foundation
Science & Technology (no. 10CG51).
References
[1] Babrauskas V. Performance-based fire safety engineering design: the role of
fire models and fire tests. Interflam 1999. In: Proceedings of the 8th
conference. Edinburgh, Scotland; 1999. p. 799–807.
[2] Meacham BJ. Risk-informed performance-based approach to building regula-
tion. In: Proceedings of the 7th international conference on performance-
based codes and fire safety design methods. Bethesda, MD; 2008. p. 3–14.
[3] Hall JR, Watts JM. Fire risk analysis. Fire protection handbook. 20th edition.
NFPA; 2008. (Chapter 3–8), (3-135-3-143).
[4] Sekizawa A. Fire risk analysis: its validity and potential for application in fire
safety. In: Gottuk D, Lattimer B, editors. Proceedings of the 8th international
symposium on fire safety science. Beijing, China; 2005. p. 85–100.
[5] Meacham BJ. Understanding risk: quantification, perceptions, and character-
ization. Journal of Fire Protection Engineering 2004;14:199–227.
[6] He YP, Wang J, Wu ZK, Hu L, Xiong Y, Fan WC. Smoke venting and fire safety
in an industrial warehouse. Fire Safety Journal 2002;37:191–215.
[7] Hietaniemi J. Probabilistic simulation of fire endurance of a wooden beam.
Structural Safety 2007;29:322–36.
[8] Johansson H, Malmnäs PE. Application of supersoft decision theory in fire risk
assessment. Journal of Fire Protection Engineering 2004;14:55–84.
[9] Wong LT, Cheung TF. Evaluating probable risk of evacuees in institutional
buildings. Safety Science 2006;44:169–81.
[10] Tanaka T, Yamaguchi JI. A consideration on determination of design fire
based on fire risk concept. Fire Science and Technology 2006;25:115–32.
[11] Upadhyay RR, Ezekoye OA. Treatment of design fire uncertainty using
Quadrature Method of Moments. Fire Safety Journal 2008;43:127–39.
[12] Hanea D, Ale B. Risk of human fatality in building fires: a decision tool using
Bayesian networks. Fire Safety Journal 2009;44:704–10.
[13] Oberkampf WL, DeLand SM, Rutherford BM, Diegert KV, Alvin KF. Error and
uncertainty in modeling and simulation. Reliability Engineering and System
Safety 2002;75:333–57.
[14] Bwalya A. An overview of design fires for building compartments. Fire
Technology 2008;44:167–84.
[15] NFPA. NFPA 204 guide for smoke and heat venting-1998 edition. National
Fire Protection Association; 1998.
[16] Holborn PG, Nolan PF, Golt J. An analysis of fire sizes, fire growth rates and
times between events using data from fire investigations. Fire Safety Journal
2004;39:481–524.
[17] Sekizawa A, Ebihara M, Notake H. Development of seismic-induced fire risk
assessment method for a building. In: Evans D, editor. Proceedings of the 7th
international symposium on fire safety science. Worcester, MA, USA; 2002.
p. 309–320.
[18] Evans DD, Stroup DW. Methods to calculate the response time of heat and
smoke detectors installed below large unobstructed ceilings. Fire Technology
1985;22:54–65.
[19] Bucci P, Kirschenbaum J, Mangan LA, Aldemir T, Smith C, Wood T. Construc-
tion of event-tree/fault-tree models from a Markov approach to dynamic
system reliability. Reliability Engineering and System Safety 2008;93:
1616–27.