SMOKE DETECTION PERFORMANCE AND FDS

MODELLING FOR FULL-SCALE FIRE TESTS

Yun Jiang1
Xtralis Pty Ltd
Bentleigh East, VIC 3165, Australia
Peter Massingberd-Mundy
Xtralis (UK) Ltd

ABSTRACT
Fire engineering designs require quantified performance of fire detection systems to
comply with deemed to satisfy (DtS) or performance based design (PBD)
requirements. Nominal sensitivities/thresholds specified by the manufacturers are
usually adopted to determine the detection times in various calculations and
computer modellings. Challenges are continuing raised on certainty of detectors
performance in real fires, and accuracy of computer modelling of the fire detection
systems. As a principle participator in the Firegrid research project, Xtralis undertook
full-scale fire tests of the real performances of various fire detection systems,
including conventional point type detectors and Air-sampling (aspirating) Smoke
Detection (ASD) systems. Furthermore, validation of computer modelling on the
detection was achieved from the in-situ tests.
This paper presents major test and simulation results of fire detection from two sets
of full-scale fire tests conducted in the Firegrid project. It reveals performance
differences between the various detection systems studied and also the accuracy of
predictions generated using computational fluid dynamics (CFD) models. The results
show that Fire Dynamics Simulator (FDS) simulation on the ASD system generated
quite high accuracy of performance prediction in those tested scenarios, while
significant inconsistencies exist between the experimental and simulation results for
the conventional point type smoke detectors. In the tests the smoke levels observed
at the time of alarm of the photoelectrical detectors exceeded the manufacturers
nominal sensitivities by more than 220 to 580%. The simulated average alarm times
of the point detectors are between 30 to 56% of the tested. These inconsistencies are
believed to be attributed to a combination of the uncertainty of point detector
response in complex fire environments and an over-estimated soot level developed in
big fires by the FDS simulations.

INTRODUCTION
Point type smoke detectors, including photoelectrical (optical) and ionisation types,
are fire detectors widely adopted in commercial and residential use. However their
detection performance is frequently questioned when quantified detection
performance is required in fire safety engineering designs. While their basic response
to different fires (selective detection) is well understood, (ionisation detectors being
more responsive to flaming fires and photoelectrical detectors to smouldering fires), a
major issue remains as to the predictability of their response in real fire situations.
These issues have been highlighted in a number of recent researches[1] [2], and the
reported errors relative to detectors nominal sensitivities published can be high as
1000%[3]. Such inconsistencies occur particularly under complex combustion and
environmental (ventilation) conditions.
1

Dr. Yun Jiang, Senior Applications Engineer, Xtralis ([email protected])

Y JIANG / P MASSINGBERD-MUNDY

1 / 16

SMOKE DETECTION PERFORMANCE AND FDS MODELLING FOR FULL-SCALE FIRE TESTS

ASD systems are increasingly being deployed in the protection of commercial and
residential buildings. These systems actively draw samples from the protected area
and, as such it is both logical and has been demonstrated that predictions of their
response using CFD are significantly more accurate than with point type detectors.
During the past 2 decades, ASD systems have been used in many challenging
applications, such as large open spaces (LOS), high velocity environments
(Cleanroom) and very early warning (VEW) detection required environment (IT&C).
More recently, the high accuracy in quantified detection performance has seen an
increasing use of ASD in performance-based design (PBD) solutions implemented by
many fire consultants and engineers.
Computer fire modelling is becoming a powerful and useful tool in fire safety
engineering designs, especially in PBDs where qualitative and quantitative
performances are prerequisite. Great progress has been achieved recently on using
computer models, especially CFD models to predict the performances of fire
detection systems. Most significantly, FDS version 5 now integrates sub-models for
various detection technologies, such as point, beam-projected and ASD type smoke
detectors[4]. Researchers suggest that a higher accuracy prediction for point
detectors activation performance can be expected by modelling real-time flow entry
behaviours of such detectors[5]. However, accuracy in smoke detection remains
relative poor from those sub-models compared to predictions for other, more
fundamental phenomena, such as temperature and flow velocity, etc. Therefore,
more validations for a wider range of detectors and fire scenarios by laboratory and
in-situ tests are required to improve the models and achieve a higher grade of
confidence.
Firegrid is a multi-million pound, multi-participator fire research project supported by
the UK government. Firegrid aims to use various formats of data from sensor
networks combined with grid and high performance computing in order to deliver realtime modelling data to an Artificial Intelligent (AI) based command station to provide
prediction of fire development and structural response to fires. As such the project
provided a great opportunity to test detection technologies under various full-scale
fire test conditions. Furthermore, due to the richness of data gathering, CFD
modelling results can be validated against experimental test data using a number of
variables such as temperature, radiative flux, flow velocity, deflection, gas
concentration as well as smoke level.

RESEARCH OBJECTIVE AND APPROACH

In line with the general objective of Firegrid, to provide fire detection data to the grid
for real-time modelling and prediction, Xtralis set its own objectives for the full-scale
tests as follows:
Obtain actual performance of ASD and conventional fire detectors from those
real fire scenarios;
Conduct FDS modelling of fire development and detection performance;
Validate the prediction of smoke detection from various modelled and criteria
with the experiments.

2 / 16

Y JIANG / P MASSINGBERD-MUNDY

SMOKE DETECTION PERFORMANCE AND FDS MODELLING FOR FULL-SCALE FIRE TESTS

To achieve the above objectives, the following approach was adopted in the
research:
Use a high sensitivity ASD system to accurately monitor and record smoke
levels adjacent to point detectors, during alarm activation of detection
technologies;
Validate the simulated ASD detection performance by the test data;
Evaluate the alarm prediction accuracy of various computer modelling
methods for point type detectors by both actual point detector and the ASD
(reference) readings.

FULL-SCALE IN-SITU TESTS

Two sets of full-scale fire tests were conducted in a high-rise apartment in
Dalmarnock, Glasgow (July 2006) and in a test rig built in Building Research
Establishment (BRE), London, (Oct 2008) respectively.
1.
DALMARNOCK APARTMENT TESTS
Test Rooms
The tests were conducted in two sets of apartments, in Level Four and Six of high
rise blocks, having identical layouts. Each of the apartments consisted of a lounge, a
kitchen, two bedrooms and a bathroom, as shown in Figure 1. The fire started from
the Lounge, which has dimensions of 6.4 x 4.0 x 2.7 metres (m) and was the only
furnished room, as illustrated in Figure 2.

Figure 1: Floor Plan of the Test Apartment

Y JIANG / P MASSINGBERD-MUNDY

3 / 16

SMOKE DETECTION PERFORMANCE AND FDS MODELLING FOR FULL-SCALE FIRE TESTS

(a) Settings in Lounge

Figure 2: Room of Fire Origin and Pilot Source

(b) Pilot Source

Fire Source and Test Conditions

Two fire tests: named controlled (ventilation) and un-controlled, were performed in
the two apartments, with identical configurations in the fire source room (Lounge) and
ignition condition. The un-controlled fire test results are presented in this paper to
represent more realistic fire condition.
Ignition of the fire was from a plastic waste paper bin filled with crumpled newspaper
and 500ml of heptane. A blanket was draped on one arm of the sofa to help flame
spreading to the sofa, as shown in Figure 2 (b).
During the un-controlled fire test, all doors were kept open, except the Bath, Storage,
and the door connecting Bedroom2 and Balcony. All windows were closed except
those in Kitchen (ajar) and in Bedroom2 (open with fly screen).
Fire Detectors
An ASD and three point type detectors, ionisation, optical-smoke and heat detectors,
were installed in each major room of the apartment. Their locations are indicated in
Figure 1 in as Detection Stations (DS). Each DS contained three point type detectors
and one sampling hole of the ASD system. A total of five ASD detectors, one
sampling hole per detector, were used to monitor absolute smoke levels (without
dilution) at each DS.
The Alert and Fire alarm thresholds of the ASD system were set to 0.05 and
0.20%OBS/m. The Fire alarm threshold of the point type smoke detectors (ionisation
and optical) was set to 2.1%OBS/m. The Fire alarm threshold of the heat detectors
was set to 58oC.
2.
BRE TESTS
Test Rig
The test rig in BREs Burn Hall consists of three rooms of identical size, 2.4 x 2.4 x
2.4 m. One corridor, with dimensions of 3.6 x 1.2 x 2.4 m, in the middle connects all
of the rooms, shown in Figure 3. All doors are the same size as 2.0 x 0.8 m.

4 / 16

Y JIANG / P MASSINGBERD-MUNDY

SMOKE DETECTION PERFORMANCE AND FDS MODELLING FOR FULL-SCALE FIRE TESTS

Rooms 1 & 4 composed of identical wall materials (double plaster boards with
100mm Rockwool in middle) and ceiling materials (Supalux-inside and plaster boardoutside sandwiching with 100mm Rockwool). Rooms 2 & 4 comprised single plaster
board on timber studs for walls and ceiling. The floor in all rooms is made of
concrete. Sampling locations of partial measured variables are marked in Figure 3 as
well.

Figure 3: Floor Plan of the Test Rig in BRE

Fire Source and Test Conditions

Various fire scenarios were tested in the test rig, including: smouldering timber stick,
smouldering cotton braid/wick, overheated PVC wire, medium-size propane gas
burner combustion, and finally a full load fire using a sofa as the main fuel load. This
paper presents results from the big fire test only. This fire was piloted from the
smouldering of cotton braid which after 2-3 minutes ignited a match (to provide a
definitive point of ignition). The match ignited 4 small pieces of fibre boards (approx.
80 x 25 x 12 mm) soaked in industrial methylated spirits to spread the flame to the
sofa. The configuration of the fire room (Room1) and the piloted source are pictured
in Figure 4. During the test, all doors were kept open except the external door to
Room1 which was for access during the setup only.

(a) Settings in Room1

Figure 4: Room of Fire Origin and Pilot Source
Y JIANG / P MASSINGBERD-MUNDY

(b) Pilot Source

5 / 16

SMOKE DETECTION PERFORMANCE AND FDS MODELLING FOR FULL-SCALE FIRE TESTS

Fire Detectors
The same types of detectors as those used in the Dalmarnock apartment test were
adopted in this test. Detection Stations (DS) were placed under ceiling in the middle
of each room, as indicated in Figure 3.
Sensitivities of the smoke detectors, including the ASD and the point type smoke
detectors, were set as those in the Dalmarnock experiment. The Fire alarm threshold
of the heat detectors were changed to 61oC.

FDS MODELLING SETTINGS

The FDS versions 4.0.7 and 5.2.0 were utilised to conduct the computer modelling of
each test. Settings of the models for the two tests are described below.
1.
SIMULATION SETTINGS FOR DALMARNOCH TEST
Multiple meshes were introduced in the simulation. In the fire room (Lounge), the grid
resolution was 0.05 x 0.05 x 0.05 m. In the rest of the apartment, 0.1 x 0.1 x 0.1 m
resolution was applied. Illustration of the domain is shown in Figure 5.

Figure 5: Simulation Domain and Meshes

The combustion behaviours of the sofa in the Lounge was defined in the simulations
by two methods: 1) a generic HRR profile from a NIST sofa burning test[6], and 2) a
HRR profile based on the average values of a Round-Robin simulation for the tests[7].
These HRR curves are illustrated in Figure 6.

6 / 16

Y JIANG / P MASSINGBERD-MUNDY

SMOKE DETECTION PERFORMANCE AND FDS MODELLING FOR FULL-SCALE FIRE TESTS

700
Sofa-RR

600

Sofa-NIST

HRR (kW)

500
400
300
200
100
0
0

200

400

600

800

1000

Time (s)

Figure 6: HRR Curves Adopted in the Simulations

2.
SIMULATION SETTINGS FOR BRE TEST
A higher grid resolution, 0.05 x 0.05 x 0.05 m, was applied for the fire room (Room1)
and the corridor (Room4), and 0.1 x 0.1 x 0.1 m resolution was applied to the other
rooms. Figure 7 illustrated the simulation domain.
From a free burning test before the formal test, it was established that the HRR
profile from the sofa burning is similar to a T2 slow growth trend with a peak value of
400kW. As such, a T2 slow growth profile was specified to the sofa in the simulation
preceded by a 3-minute incipient stage (30W constant fire).
One particular feature in the test was the positioning of an additional light in the left
hand bottom corner of Room1. This 500W halogen lamp was needed for the
cameras, and influenced the flow pattern and smoke movement during the incipient
stage. This lamp was simulated as a convective heat source with estimated 300W
output.

Figure 7: Simulation Domain

SMOKE DETECTION SIMULATED IN FDS

1.
ASD DETECTOR
As described by the author previously[8], smoke density in an ASD central detection
unit is a function of smoke densities and flow rates at each sampling hole of the
Y JIANG / P MASSINGBERD-MUNDY

7 / 16

SMOKE DETECTION PERFORMANCE AND FDS MODELLING FOR FULL-SCALE FIRE TESTS

detector as well as transport times from the sampling holes to the central unit. While
the smoke density distribution can be simulated by FDS, the other parameters (flow
rate and transport time) need to be calculated separately. Most usually, this is done
by using special pipe network computing software developed by one of the major
ASD manufacturers. For this paper, the parameters of the ASD system were
calculated by ASPIRE2, a software developed by Xtralis.
FDS version 5 now supports calculation of accumulation of smoke from multiple
sampling holes in an ASD system once the flow and transport time for each individual
sampling hole is provided. However, the data presented in this paper for the
Dalmarnock test was performed using FDS version 4. As such the calculation was
calculated manually. For validation, the same manual calculations were also done on
the simulations generated using FDS5 and compared with the results obtained using
the new automatic accumulation code in FDS5 - identical results were obtained.
2.
POINT TYPE SMOKE DETECTORS
Three methodologies for predicting the point detectors response were adopted:
1) using built-in smoke detector sub-models from FDS,
2) applying averaged optical densities from statistics, and
3) by multiple criteria.
Two types of smoke detection sub-models, Cleary and Heskestad models, have been
built in FDS to simulate the response of point type smoke detectors[4]. The behaviour
of smoke entering into the detection chamber was simulated by various numbers of
parameters. In Heskestads model, a single parameter model, dwell time was
expressed by time for smoke travelling over a characteristic length L. In Clearys
multi-parameter model, the smoke entry lag time is expressed as sum of dwell time
and mixing time, which are computed by two parameters respectively. NIST
developed a test facility, FE/DE to determine those required parameters for point type
smoke detectors[9]. However, without enough validation results, applicability and
suitability of those parameters in real fire scenarios are still unclear since
environmental conditions surrounding the detectors may be quite different from the
standard test. Furthermore, such parameters are still unavailable for most of todays
point type smoke detectors, even for widely used ones such as those used in the
Firegrid tests. Therefore, the default parameters for the Cleary model suggested by
the FDS manual were adopted in the current simulation. As for the Heskestad model,
characterised length values as 1.8m for the ionisation detector and 15m for the
optical detector, recommended from SFPE handbook[10], were used.
The second method is to use an average optical density (OD), rather than the
detectors nominal sensitivity, to achieve a higher certainty in detection. As
recommended by Geiman and Gottuk[1], optical density of 0.11 OD/m 0.04 as the
primary indicator of a smoke alarm can result in 80% certainty from various research
tests (i.e. 80% of detectors tested have responded by this OD). This method has
been used in number of studies, including fire modelling conducted for a Fire
Protection Research Foundation (NFPRF) Detector Performance In Waffle Ceilings
research project[11]. To assess the ionisation and photoelectrical detectors used in the
current study individually, values suggested by the author for this two types of
detectors were adopted as: ionisation detector, 0.072 OD/m; photoelectrical detector,
0.11 OD/m. While the value for photoelectrical detectors is applicable for both flaming
8 / 16

Y JIANG / P MASSINGBERD-MUNDY

SMOKE DETECTION PERFORMANCE AND FDS MODELLING FOR FULL-SCALE FIRE TESTS

and smouldering fires, the value of 0.072 OD/m is just for flaming fires. However,
since the fire was flaming in both tests before either type or detector operated it is
reasonable to use the latter figure for the ionisation detector.
The third method in determining the alarm status of the point type detectors is by
satisfying certain criteria, including optical density, temperature rise and gas velocity
at the postulated smoke detector locations. Thresholds of those criteria were given as
below:
a) The optical density was decided by the detectors nominal sensitivity.
b) Temperature rise (T) was 4 oC for ionisation detectors and 13 oC for
photoelectric detector (representing conservative thresholds).
c) Velocities should be above 0.13 m/s 0.06.

RESULTS AND DISCUSSION

1.
RESPONSE TIMES OF THE DETECTORS FROM THE TESTS
The detectors response times from the two tests are shown in Table 1 and Table 2.
Since an incipient stage burning was included in Test2, ASDs dual stage response
times, at Alert and Fire alarms, were given in the tables.
Table1DetectorsResponseTimesfromTest1
Location
Lounge
Kitchen
Hall (Corridor)
Bedroom1
Bedroom2

Alarm Response Time (s)

Ionisation
Optical
ASD
Alert Alarm Fire Alarm Fire Alarm Fire Alarm
9
9
16
25
14
17
30
57
23
23
84
83
42
42
98
100
227
227
234
234

Heat
Fire Alarm
40
230
109
307
328

Table2DetectorsResponseTimesfromTest2
Location
Room1
Room2
Room3
Room4

Alarm Response Time (s)

ASD
Ionisation
Optical
Alert alarm Fire alarm Fire alarm Fire alarm
78
98
200
226
117
252
273
313
100
239
247
248
71
79
230
261

Heat
Fire alarm
440
863
813
702

In the two tests, ASD had earlier response to smoke than the point detectors, in all
locations. When an incipient stage is presented, an ASD system can provide very
early warning (VEW) fire detection to small amounts of smoke generated from a
smouldering burning, as shown from the 2nd test. ASDs early detection feature is
attributed to a combination of its high sensitivity and stable and reliable performance
from its active sampling technique. The advantages of very early warning (VEW)
are clearly reduced where there is no incipient stage in the development of the fire
such as in Test1 when a significant flaming fire was set at the beginning of the test.
Its interesting to note that in Test2, the ASD detector in the corridor (Room4) raised
an alarm earlier than that in the fire room (Room1). This unexpected result can be
attributed to the 500W lamp installed in the corner of the fire room. Built-up of
convective flows in the room caused by hot air rising from this lamp meant that
smoke rising from the smouldering cotton braid drifted out of the fire room through
Y JIANG / P MASSINGBERD-MUNDY

9 / 16

SMOKE DETECTION PERFORMANCE AND FDS MODELLING FOR FULL-SCALE FIRE TESTS

the open door during the incipient burning stage away from the sampling hole in
the middle of Room1. When the fire developed into a flaming fire, the convective
airflows in the room became dominated by the fire rather than the lamp, and as such,
the situation was totally different by the time the fire was bigger enough to trigger the
conventional smoke detectors. Hence as one might expect, the detectors in the fire
room responded to the fire quicker than those in the corridor.
2.

ACTUAL SMOKE LEVELS AT POINT DETECTORS RESPONDED

Figure 8 shows smoke levels (in percentage) recorded by the ASD detectors at
detection times, appropriate transport times applied for corrections, of those point
type smoke detectors. The nominal sensitivities of the point detectors are 2.1%/m.
1523
Ion

800

O pt
Ion
600

O pt

400

200

0
L ounge K itc hen

Hall

Test1

B ed1

B ed2

R m1

R m2

R m3

R m4

Test2

Figure 8: Percentage of Monitored Smoke Levels to the Nominal Sensitivity

In Test1, the ranges of smoke levels recorded for the ionisation and optical detectors
are from 81.0% to 1523.8% and from 171.0% to 1523.8% of the nominal sensitivity.
In Test2, the ranges are from 11.9% to 35.7% and from 27.6% to 328.6% for the
ionisation and optical detectors respectively. Generally higher smoke levels were
recorded for the both types of detectors in Test1, where a fire developed quickly and
complicated geometry conditions were presented.
The ionisation detectors response appears significantly different in the two test fires
when assessed against the smoke density measured using the ASD. In contrast the
optical detectors seem to show a more consistent response. This is not particularly
surprising because the ASD system used in the comparison incorporates optical
laser technology to determine the smoke density. While the optical points used
incorporate a similar optical (but LED based) technology, the ionisation detectors
incorporate a fundamentally different detection technology. Clearly the ionisation
detectors were relatively more responsive to the products of combustion produced in
the early phases of Test2 than in Test1. What is worth noting is that the nominal
sensitivities used are essentially related to optical density, not to an ionisation
reference. As such it is not unexpected that the ionisation detectors may respond
much faster than is predicted by a measurement of optical density particularly
where a flaming fire is in progress. In some research this tendency to generate
alarms much lower than its nominal sensitivity has been described as the selective
detection feature of ionisation detectors. Recent research[12], recorded smoke levels

10 / 16

Y JIANG / P MASSINGBERD-MUNDY

SMOKE DETECTION PERFORMANCE AND FDS MODELLING FOR FULL-SCALE FIRE TESTS

at ionisation detectors (with sensitivities between 2.5 to 5 %/ft) activated are below
than 1%/ft.
The optical detectors respond at smoke levels much higher than their nominal
sensitivity in the two tests consistently. The one exception to this observation is the
detector in Room3 in Test2. Response of this detector is questionable since its alarm
time is earlier than the one in Room4 which is closer to the fire room. Excluding the
result from Detector 3 in Test2, the mean values of the recorded smoke levels at
alarm for the optical detectors in the two fires are 5.8 and 2.2 times of the nominal
sensitivity respectively. These results match with other researches[2][12].
SIMULATED DETECTORS PERFORMANCES
3.
The ASD Detector
From the simulations for Test1, higher smoke density was obtained from simulations
conducted using FDS5 compared with using FDS4, and the discrepancy increased
as the fire developed. However, simulated ASD detection performances from these
two versions are quite close since the high sensitivity ASD detectors are capable of
responding to very low smoke density occurring during the early stage of the
combustion, when the discrepancies between FDS4 and 5 were small. Therefore, for
Test1, this paper presents results from simulations by version 4. Results obtained
from the two HRR setting methods, from the NIST sofa burning and the RR
simulations, were presented in Table 3 along with the actual response times
measured during the tests.
Table 3: Simulated ASD Response Times Test1
Tested
Percentage of Resonse
Time (s)
Tested (%)
111.1
9
111.8
17
95.7
23
97.6
42
22.9
227

FDS Simulated Response Time (s)

Detector
Location

NIST

Lounge
Kitchen
Hall
Bedroom 1
Bedroom 2

9
18
21
35
46

Percentage of
Tested (%)
100.0
105.9
91.3
83.3
20.3

RR
10
19
22
41
52

No significant difference exists between these two setting methods of the fire source
even through the HRR profiles are different. Furthermore, the predicted response
times were all within 20% of the times recorded during the test with the one
exception of Bedroom2. This is because the open windows in Bedroom2 were not
included in the simulation. From video footage of the experiment it was observed
that during the test, smoke spread into Bedroom2 was postponed by wind entering
into the apartment through the external hall and a fly screen mounted on top of the
Bedroom2 windows. When the smoke front overcame the resistance and eventually
entered into Bedroom2, smoke density became quite high which resulted similar
response times from the ASD and point type detectors, as shown in Table 1. Such
wind effect was not simulated in the modelling and significant errors occurred in
simulated response times not only for the ASD detector but also for the point
detectors in the Bedroom2.
Predictions of the ASD detector response times in Test2 are given in Table 4 based
on Alert alarm set at 0.05%/m and Fire alarm at 0.2%/m. Also included in the table
are the comparisons to the time measured during the fire.
Y JIANG / P MASSINGBERD-MUNDY

11 / 16

SMOKE DETECTION PERFORMANCE AND FDS MODELLING FOR FULL-SCALE FIRE TESTS

Table 4: Simulated ASD Response Times Test2

Location
Alarm Level
Response Time
(s)
Response Time
from Test (s)

Room1
Alert Fire

Room2
Alert Fire

Room3
Alert Fire

Room4
Alert
Fire

83

92

140

177

112

130

73

81

78

98

117

252

100

239

71

79

Percentage of the
106.4 93.9 119.7 70.2 112.0 54.4
tested (%)

102.8 102.5

As can be seen, the predictions, particularly during the incipient stage of the fire are
reasonably accurate. The predictions of the Alert alarm times, corresponding to the
first hints of smoke at the sampling locations, were dramatically affected by modelling
of the flow pattern. In simulations which included the 500W lamp, the built-up of
convective flow form the lamp was reasonably simulated in FDS, as illustrated in
Figure 9. However, simulations performed prior to the live test were performed with
the 500W lamp excluded and predicted that the ASD detector in Room1 would alarm
first and that responses of the ASD detectors in other rooms would also be slower
than the values presented in Table 4.
This clearly demonstrates the importance of considering and including all possible
heat and ventilation sources in any simulations of incipient conditions as it is the
natural (or residing) air flows which determine where smoke will travel and how early
warning and very early warning detectors will respond.

(a) Elevation View at Centreline

(b) Top View at 2m High
Figure 9: Simulated Flow Pattern (Velocity) in Fire Room Before Flaming

Considering the response predictions further into the fire (after the end of the
incipient phase and some time after ignition) it is clear that the flows become
dominated by the heat generated by the fire. However, the smoke generation model
used in FDS is also a significant parameter in determining the accuracy of the
response prediction. As shown in Figure 10, simulated smoke densities in all of the 4
rooms, expressed as obscuration levels, are much higher than those recorded during
the test. (Its noted that the maximum value can be displayed by the ASD detector is

12 / 16

Y JIANG / P MASSINGBERD-MUNDY

SMOKE DETECTION PERFORMANCE AND FDS MODELLING FOR FULL-SCALE FIRE TESTS

32%/m.) Such over-predicted smoke concentrations by FDS has been reported by

other researchers[13]. There are two principle reasons for this over-prediction;
1) Firstly, in modelling of soot generation. Soot yield adopted in FDS simulations
are usually obtained from small scale tests[14]. It is reported that as the fire size
increases the soot yields decreases[15]. Difference between measured soot
yields from tests and those reported in literatures ranged from a factor of 2 to
5[13]. Therefore much higher smoke density was predicted by FDS under large
fire size conditions comparing to the test results.
2) Secondly, in modelling of soot transportation. Soot deposition on ceiling above
the fire source in the test can significantly reduce amount of smoke
transported to sampling locations. According to a test conducted under a 2.7m
high ceiling, 39% of soot was removed from combustion of a 100kW propylene
fire in the first 5 minutes, by depositing on the ceiling[13].

Obscuration (%/m)

100

Room1-Simulated

90

Room2-Simulated

80

Room3-Simulated
Room4-Simulated

70

Room1-Tested
Room2-Tested

60

Room3-Tested
Room4-Tested

50
40
30
20
10
0
0

100

200

300

400

500

600

Time (s)

Figure 10: Smoke Obscuration from the Simulation and Test Test2

The Point Type Smoke Detectors

Predicted response times for the point type smoke detectors in Test 1&2 are shown
in Table 5 and Table 6 respectively along with comparisons with the times recorded
during the tests. The predicted response times for Test1 were generated by Cleary
model with the default parameters, while those for Test2 were generated using 4
different methods (as indicated in the table).
Table 5: Simulated Response Times for the Point Detectors Test1
Detector
Location
Lounge
Kitchen
Hall
Bedroom
1
Bedroom
2

Detector
Type
Ionization
Optical
Ionization
Optical
Ionization
Optical
Ionization
Optical
Ionization
Optical

Tested
FDS Simulated Response Time (s)
Percentage
Percentage Resonse
NIST
RR
of Tested (%)
of Tested (%) Time (s)
9
56.3
7
43.8
16
36.0
28.0
9
7
25
100.0
70.0
30
21
30
54.4
35.1
31
20
57
33.3
23.8
28
20
84
33.7
24.1
28
20
83
55
56.1
40
40.8
98
54.0
39.0
54
39
100
35.9
21.8
84
51
234
35.9
21.8
84
51
234

Table 6: Simulated Response Times by Various Methods Test2

Y JIANG / P MASSINGBERD-MUNDY

13 / 16

SMOKE DETECTION PERFORMANCE AND FDS MODELLING FOR FULL-SCALE FIRE TESTS
Room1
Room2
Room3
Room4
Ionisation Optical Ionisation Optical Ionisation Optical Ionisation Optical
Cleary
model

Response time
(s)

Percentage. of
the tested (%)
Response time
Heskestad (s)
model
Percentage. of
the tested (%)
Response time
Average
(s)
OD
Percentage. of
method
the tested (%)
Response time
(s)
Multicriteria Percentage. of
the tested (%)

204

204

246

244

231

229

207

206

102.0

90.3

90.1

78.0

93.5

92.3

90.0

78.9

199

201

235

246

226

233

203

210

99.5

88.9

86.1

78.6

91.5

94.0

88.3

80.5

202

210

251

257

238

244

224

228

101.0

92.9

91.9

82.1

96.4

98.4

97.4

87.4

201

214

251

269

238

253

219

243

100.5

94.7

91.9

85.9

96.4

102.0

95.2

93.1

From Table 5 it is clear that the FDS simulations over-predict the detector
performance i.e. that they will detect earlier than they actually do during the test.
The simulated response times were 20 to 50% of the tested result in most cases. The
predictions generated using the HRR profile created by the Round-Robin simulations,
which has lower peak value and is believed closer to the real combustion condition
compared to that from the NIST sofa burning profile, actually resulted in even early
and more misleading predictions. The shorter response times from the RR
simulation, 40.0% and 29.6% of the nominal sensitivity for the ionisation and optical
detectors compared to 56.3% and 42.8% accordingly from the NIST simulation, may
be caused by the 1st quick rising peak in the HRR curve. As per the discussion
above with respect to the effect of over-predicted soot levels by FDS the soot
generation at this peak (approximated 160kW) calculated by the simulation would be
much higher than would actually occur in the real fire burning. Such uncertainty
should be taken into consideration when complex geometry, boundary and
combustion conditions are modelled.
In the relatively simple test condition (Test2), where most of detection stations
located in the centerline of the open doors and were closer to the fire source, much
higher accuracy of the predicted detection performance has been achieved for the
point detectors. Generally, the predictions for the ionisation detectors are more
accurate compared to the photoelectrical detector, due to a fact that flaming fire from
the liquid fuels existed in early stage in both tests and that favors the ionisation
detector.
The single parameter Heskestad model generated a similar accuracy as the multiple
parameters Cleary model, by using the default and normal suggested values for the
parameters. This matches with others observation[5]. However, comparing to Cleary
model, Heskestad model produced a better trend to match that tested, i.e. slower
responses from the optical detectors than the ionisation detectors at the same
location. This indicates that Cleary model much more over-predicted detection
performance for the optical detectors.
By applying the average OD method and the multi-criteria method, prediction of the
response times for all the point type detectors has been improved. Among all the
methods investigated in the current study, prediction by the multi-criteria method
produces highest accuracy. The improvement between predictions by the multicriteria method and the Cleary model can reach to 14% in Test2.
14 / 16

Y JIANG / P MASSINGBERD-MUNDY

SMOKE DETECTION PERFORMANCE AND FDS MODELLING FOR FULL-SCALE FIRE TESTS

Generally, over-predicted soot density from FDS simulation has greater impact on
prediction accuracy of detection performance of point type detectors due to their
relatively lower sensitivity and passive detection feature. Its under certain level of
risks to predict point type detectors performance based on their nominal sensitivities,
even by FDS built-in detection models. Utilizing the average OD method or multicriteria method can improve prediction certainty of the point type detectors, especially
for the optical detectors.

CONCLUSIONS
The ASD system during the in-situ fire tests exhibited very early response compared
to the normal sensitivity point detectors (ionisation and optical detectors). The ASD
very early response capability was consistent across all fire scenarios from large
HRR to incipient (smouldering) fires (with very low heat and smoke product
generations).
The ASD system detection performance has been simulated successfully by FDS
providing reasonable accuracy in prediction. Therefore when modelled ASD
performance is to be applied into FSE designs, including PBD, relative higher degree
of confidence is expected, compared to conventional detectors.
The performance of the tested point type detectors during the in-situ fire tests was
unstable. Higher smoke levels than their nominal sensitivities, with a factor of 1.5 to
5, were required for the photoelectrical detector. Response of the ionisation detector
is somehow hard to predict and depending on fuel type of combustion greatly, from
20% to 1500% of its nominal sensitivity. How accurately simulated the combustion
profile will affect accuracy of performance prediction on point type smoke detectors.
Higher soot densities simulated in FDS, especially in version 5, usually over-predict
the performance of the point type smoke detectors.
Issues existing in modelling of soot generation and transportation, and in modelling of
entry behaviors of point type detectors may introduce significant errors and
uncertainty in fire detection in fire engineering designs by applying the detectors.
Certain degree of conservative should be taken into design practices. Investigated
average OD method and multi-criteria method can improve accuracy in prediction of
detection performance for the point detectors.
Future analysis of full set of experimental data and study soot yields under various
combustion conditions will increase understanding of fire behaviours and further
improve FDS modelling on fire products and then fire detection.

REFERENCES
[1]
[2]
[3]

Geiman, J.A. and Gottuk D.T., 2002, Alarm Thresholds for Smoke Detector
Modelling, Fire Safety Science Proceedings of the Seventh International
Symposium
Geiman, J., Gottuk, D., and Milke, J., 2004, Evaluation of Smoke Detector Response
Estimation Methods, 8th NFPA Fire Suppression & Detection Symposium
Geiman, J.A., 2003, Evaluation of Smoke Detector Response Estimation Methods,
Thesis of Master degree, University of Maryland

Y JIANG / P MASSINGBERD-MUNDY

15 / 16

SMOKE DETECTION PERFORMANCE AND FDS MODELLING FOR FULL-SCALE FIRE TESTS

[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]

[12]
[13]
[14]
[15]

McGrattan, K.B., Klein, B., Hostikka, S., et. al, Forney, G.P., 2007, Fire Dynamics
Simulator (Version 5) Users Guide, NIST Special Publication 1019-5, National
Institute of Standards and Technology, Gaithersburg, MD
Roby, R.J., Olenick, S.M., Zhang, W., et. al, 2007, A Smoke Detector Algorithm for
Large Eddy Simulation Modelling, NIST GCR 07-911, National Institute of Standards
and Technology, Gaithersburg, MD
NIST, Fire Experiment Results, National Institute of Standards and Technology,
http://www.fire.nist.gov/fire/fires/fires3.html
G. Rein et. al, 2007, Round-Robin Study of Fire Modelling Blind-Predictions using the
Dalmarnock Fire Experiments, 5th International Seminar on Fire and Explosion
Hazards, Edinburgh, April 2007
He, M., and Jiang, Y., 2005, Use of FDS to Assess the Effectiveness of an Air
Sampling-Type Detector for Large Open Space Protection, Journal of Fire Protection
Engineering, The Society of Fire Protection Engineers
Cleary, T., Chernovsky, A., Grosshandler, W., et. al, 1999, Particulate Entry Lag in
Spot-Type Smoke Detectors, 6th International Association for Fire Safety Science,
Poitiers, France
Schifiliti, R.P., Meacham, B.J., and Custer, R.L.P., 2002, Design of Detection
Systems, The SFPE Handbook of Fire Protection Engineering (3rd ed), National Fire
Protection Association, Quincy, MA
OConnor, D., Cui, E., Klaus, M., et. al, 2006, Smoke Detector Performance for level
Ceilings with Deep Beams and Deep Beam Pocket Configurations, An Analysis
Using Computational Fluid Dynamics, The Fire Protection Research Foundation,
Quincy, MA
Mowrer, F.M., Milke, J.A., Gandhi, P., 2008, Validation of a Smoke Detection
Performance Prediction Methodology, Volume 3. Evaluation of Smoke Detector
Performance, The Fire Protection Research Foundation, Quincy, MA
Gottuk, D., Mealy, C., and Floyd, J., 2008, Smoke Transport and FDS Validation, 9th
International Association for Fire Safety Science, Karlsruhe, Germany
Tewarson, A., 2002, Generation of Heat and Chemical Compounds in Fire, The SFPE
Handbook of Fire Protection Engineering (3rd ed), National Fire Protection
Association, Quincy, MA
Pitts, W.M. and Mulholland, G.W., 2000, Improved Real-fire Measurements Having
Meaningful Uncertainty Limits, NISTIR 6588, National Institute of standards and
Technology, Gaithersburg, MD

ACKNOWLEDGEMENT
This paper is based on work undertaken as part of the FireGrid Project
(www.Firegrid.org) which is co-funded by the Technology Strategy Board's
Collaborative Research and Development programme, following an open
Competition. Xtralis is a key sponsor of this project.

16 / 16

Y JIANG / P MASSINGBERD-MUNDY