Professional Documents
Culture Documents
J Firesaf 2004 07 004
J Firesaf 2004 07 004
J Firesaf 2004 07 004
Abstract
The primary goal of this study is to develop fire detection algorithms for use in residential
occupancies that reduce the nuisance sensitivity and detect fires at least as fast as conventional
ionization and photoelectric detectors. An analysis is conducted using the output from
ionization, photoelectric and carbon monoxide (CO) detectors, and a thermocouple
measurement from 32 fire tests and 11 nuisance tests. Eight parameters are identified from
the data collected from the four sensors by considering the magnitude and rate of rise of the
output from each sensor. Algorithms are developed from these eight parameters using three
approaches: analyzing the value of a single parameter relative to the response of commercial
detectors given fire and nuisance sources, conducting the same analysis with two or three
parameters and conducting the same analysis with a principal component analysis (PCA) of all
eight parameters. The best fire sensitivity and nuisance immunity was observed for three
algorithms: temperature rise and CO; CO and ionization detector; and temperature rise, CO
and ionization detector.
r 2004 Elsevier Ltd. All rights reserved.
0379-7112/$ - see front matter r 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.firesaf.2004.07.004
ARTICLE IN PRESS
2 L.A. Cestari et al. / Fire Safety Journal 40 (2005) 1–28
1. Introduction
2. Overview of experiments
In this study, data from a series of 32 fire tests and 11 nuisance tests is referenced
from the Home Smoke Fire Detector Project [6]. The fire and nuisance test matrix is
shown in Table 1. The fire tests include six smoldering items of furniture, six flaming
items of furniture, six smoldering mattresses, ten flaming mattresses, and four
flaming grease fires. The fire sources are chosen to represent typical fatal residential
fire scenarios. The mattress fires are conducted in the main bedroom. The
upholstered furniture fires are conducted in the living room, and the grease fires
are conducted in the kitchen.
The nuisance test series includes nine cooking sources, one smoking source, and
one tea candle test. Some of these nuisance sources are ‘‘aggressive’’ in the sense that
significant smoke obscuration is generated and a homeowner would benefit from
being informed of such conditions as some could develop into threatening situations.
However, as included in this experimental program, no reduction in tenability was
experienced as a result of these nuisance sources.
Measurements common to both series of tests include ionization detector analog
output, photoelectric detector analog output, residential CO detector analog output,
and a temperature measurement. Calibration correlations are available for the
measurements to convert the analog output to physical properties. Ionization and
photoelectric sensor outputs are converted to extinction coefficients. CO detector
output is converted to CO concentration and the thermal sensor output is converted
to temperature.
After the raw data are converted to physical units, a data smoothing routine is
applied to the photoelectric and carbon monoxide sensor measurements. Sensor
noise is more noteworthy for the nuisance tests. To reduce these noise-induced
ARTICLE IN PRESS
L.A. Cestari et al. / Fire Safety Journal 40 (2005) 1–28 3
Table 1
Fire and nuisance tests
Fire tests
Nuisance tests
Test ID Description
Table 1 (continued )
Nuisance tests
Test ID Description
MHN32 Baking one frozen pizza for 15 min then broiling for 5 min
MHN35 Four tea candles on the range top
MHN36 Frying bacon until crisp, but eatable
For this analysis, the alarm threshold for nuisance sources is developed simply by
creating an obscuration criterion that divides the aggressive from the non-aggressive
nuisance sources. The resulting design level of nuisance immunity is thus established
as:
1. The algorithm should alarm to nuisance tests MHN 06, 14, and 20 after they
exceed the obscuration thresholds of k=0.15 and 0.50 m1 as measured by
ionization and photoelectric detector output, respectively. Conventional detectors
would alarm to all three of these tests.
2. The algorithm should not alarm to nuisance tests MHN 09, 12, 15, 16, 19, 32, 35,
and 36. Conventional detectors would alarm to MHN 06, 09, 16, 19, and 36.
The eight parameters measured during each of the tests are the four measured
variables of ionization, photoelectric, CO and temperature analog outputs and the
ARTICLE IN PRESS
L.A. Cestari et al. / Fire Safety Journal 40 (2005) 1–28 5
rate of rise of each of these variables. The maximum responses of the eight
parameters are presented in Table 2. The following insights are gained from the
analysis of individual sensor performance during the nuisance and fire tests. These
insights assist in developing optimal sensor combinations for multi-criteria alarm
algorithms.1
Based on conventional alarm thresholds, the ionization detector is more beneficial
for detecting flaming fires, while the photoelectric detector is more beneficial for
detecting smoldering fires.
The conventional ionization and photoelectric detectors are prone to nuisance
alarms.
The rate of rise of a measurement is more beneficial for detecting flaming fires and
less beneficial for detecting smoldering fires.
The CO and rate of temperature rise measurements provide a high degree of
nuisance immunity.
CO detector output provides faster detection and fewer responses to nuisance
alarms than photoelectric detector output for smoldering fires included in this
study.
Rate of temperature rise outperforms ionization detectors for flaming fires
provided that there is no closed door between the fire and detector.
The goal of a multi-sensor algorithm is to improve the nuisance immunity and fire
sensitivity as compared to that of the single-sensor thresholds. The procedure
utilized to develop a multi-sensor algorithm begins with selecting a sensor
combination based on the performance of the individual sensors. Once a sensor
combination is chosen, the sensor data are plotted for every test to assess the
performance of hypothesized algorithms relative to the following three objectives:
provides nuisance immunity,
is at least as fire-sensitive as a conventional ionization detector, and
is at least as fire-sensitive as a conventional photoelectric detector.
Table 2
Maximum observed values for the eight parameters
Ion Ion rate Photo Photo rate CO CO rate Temperature Temp rate
(m1) of rise (m1) of rise (ppm) of rise change of rise
(ms)1 (ms)1 (ppm/s) (1C) (1C/s)
Smoldering
MHN06* 0.14 0.003 1.5 0.0511 5 0.18 11 0.11
MHN09 0.14 0.0034 0.18 0.0065 15 0.23 9 0.11
MHN12 0.02 0.0006 0 0.0000 0 0 14 0.17
MHN14* 0.14 0.0022 1.91 0.045 17 0.36 14 0.17
MHN15 0 0 0 0 0 0 0
MHN16 0.06 0.0008 0.08 0.0039 4 0.18 17 0.11
MHN19 0.14 0.0032 0.47 0.0109 1 0.14 8 0.11
MHN20* 0.14 0.0031 0.04 0.0029 1 0.1 13 0.11
MHN32 0.04 0.0007 0.02 0.0015 1 0.1 20 0.17
MHN35 0 0.0002 0 0 0 8 0.11
MHN36 0.11 0.0034 0.3 0.0058 0 0 12 0.17
Smoldering
SDC01 0.08 0.0006 0.7 0.003 253 0.3 2 0.03
SDC03 0.04 0.0003 0.067 0.0005 7 0.03 1 0.02
SDC04 0.17 0.0048 0.52 0.0078 104 0.74 27 1.15
SDC06 0.19 0.005 0.88 0.008 208 1.97 64 3.1
SDC08 0.18 0.0125 0.91 0.0272 417 5.76 90 2.43
SDC11 0.21 0.0088 0.86 0.0165 438 5.31 128 4.49
SDC31 0.25 0.007 1.16 0.0236 422 0.69 157 5.13
SDC34 0.24 0.012 0.07 0.0009 33 7.07 49 1.01
SDC37 0.25 0.006 0.4 0.0084 48 2.24 59 0.94
SDC21 0.19 0.0087 0.58 0.0149 31 0.49 31 0.16
SDC23 0.36 0.0095 0.71 0.013 178 0.6 45.75 1.6
SDC27 0.06 0.0026 0.31 0.0059 50.15 2.2 120 0.2
Flaming
SDC02 0.19 0.0042 0.42 0.006 97 0.85 48 1.14
SDC05 0.13 0.0029 0.79 0.0072 303 3.47 123 2.13
SDC07 0.18 0.0048 No data No data 418 12.79 108 2.46
SDC09 0.16 0.0104 0.93 0.0429 200 6.18 8 0.48
SDC10 0.23 0.0049 0.62 0.0155 438 7.82 156 5.63
SDC14 0.18 0.0113 1.27 0.026 414 1.75 39 3.84
SDC33 0.28 0.014 0.31 0.006 44 3.55 64 0.6
SDC35 0.25 0.015 0.26 0.0044 37 3.07 61 0.5
SDC36 0.27 0.012 0.14 0.0036 87 1.61 27 0.01
SDC38 0.28 0.011 0.13 0.0023 270 0 110 2.01
SDC39 0.25 0.001 0.1 0.0031 288 0.15 102 2.17
SDC40 No data No data 0.08 0.0034 64 0.002 27 0
SDC20 0.17 0.0058 0.57 0.0076 67 1.22 31 0.09
SDC22 0.21 0.0036 0.72 0.0209 33 1.47 45.75 0.48
SDC25 0.07 0.0026 0.43 0.0116 125 3.53 120 2.33
SDC26 0.14 0.0044 0.41 0.0135 8.24 0.04 90 1.72
Grease
SDC12 0.21 0.0033 0.77 0.009 111 1.27 43 1.28
SDC13 0.21 0.0033 0.71 0.0124 182 1.21 43 1.37
ARTICLE IN PRESS
L.A. Cestari et al. / Fire Safety Journal 40 (2005) 1–28 7
Table 2 (continued )
Ion Ion rate Photo Photo rate CO CO rate Temperature Temp rate
(m1) of rise (m1) of rise (ppm) of rise change of rise
(ms)1 (ms)1 (ppm/s) (1C) (1C/s)
Note: the nuisance source tests identified with a ‘‘*’’ involve situations where notification to the building
occupant is preferrable to remedy the situation and prevent an ignition.
2
These algorithms provide fire sensitivity and nuisance immunity for the data provided from the
experimental program conducted for the Home Smoke Fire Detector Project. While these data are from a
broad range of scenarios, the scenarios are not inclusive of all possible fire or nuisance source scenarios.
8
Table 3
Time to alarm in minutes for multi-sensor algorithms
Nuisance
ARTICLE IN PRESS
MHN06* 4.3 5.5 NA 2 NA Likely NA 4.3 6 5.5 Likely
MHN09 10.2 13 NA 2.6 3.5 Possibly NA 6.1 0 13 Possibly
MHN12 NA NA NA 1.9 18.4 NA NA NA 0 0 NA
MHN14* 5.7 5.5 8.9 1.1 6.3 Likely 8 5.7 6.8 5.5 Likely
MHN15 NA NA NA 8.9 NA NA NA NA 0 0 NA
MHN16 18 17.3 NA 0.4 NA NA NA 18 0 17.3 NA
MHN19 10.1 9.2 NA 4.5 NA Possibly NA 10.1 0 9.2 Possibly
MHN20* 5 NA NA 0.3 NA Likely NA 5 0 0 Likely
MHN32 NA NA NA 2.4 2.5 NA NA NA 0 0 NA
MHN35 NA NA NA 2.6 NA NA NA NA 0 0 NA
MHN36 5.9 5.5 NA 3.9 6.3 Unlikely NA 5.9 0 5.5 Unlikely
Smoldering
SDC01 103.8 87.8 30.3 15.3 22 30.3 30.3 30.3 92.5 87.8 100.7
SDC03 NA NA NA 23.9 NA NA NA NA NA NA NA
SDC04 55.8 19.7 17.3 11.4 14.6 17.3 17.3 17.3 55.8 19.7 55.9
SDC06 100.1 57.4 55.2 47.2 52.7 55.2 55.2 55.2 100.1 57.4 100.1
SDC08 62.2 62.5 62.3 29.5 37.3 62.2 62.1 62.1 62.3 62.3 62.2
SDC11 71 70.9 54.6 27.4 40.5 54.6 54.6 54.6 70.8 70.8 71.8
SDC31 101.82 97.2 45.2 10.6 33.1 45.2 44 44 134 84 117
SDC34 65.6 NA 58.6 39.26 50.6 58.6 21.8 21.8 64 64 65.63
SDC37 26.2 23.8 29.2 21.5 26 29.2 26.5 26.46 31.56 23.53 29.56
SDC21 33.5 33.42 34.56 18.73 21.06 33.6 33.43 33.43 34.3 33.36 33.6
SDC23 74.2 25.92 28.63 24.63 26.4 28.63 28.53 28.53 81.2 25.3 79.66
SDC27 70.9 40.26 70.16 70.16 70.16 70.16 70.16 70.26 70.8 40.06 NA
Flaming
SDC02 1.4 2.1 1 1 1 1 2.1 1.4 1 1 2.3
SDC05 0.7 1.1 0.7 0.6 0.6 0.7 1.2 0.7 0.7 0.7 3.6
SDC07 1.3 No curve 2.4 1.6 2.3 1.9 1.9 1.3 2.4 2.4 1.9
SDC09 3.1 3.6 5 2.9 3.6 3.5 3.5 3.1 18 3.6 3.5
SDC10 2.1 2.5 1.3 1.2 1.3 1.3 2.5 2.1 1.3 1.3 2.5
ARTICLE IN PRESS
SDC38 11.12 4.9 0.7 0.7 0.7 0.7 1.16 1.16 0.76 1.36 5.56
SDC39 3.8 NA 0.83 0.66 0.73 0.83 1.93 1.93 0.83 0.22 1.93
SDC40 No curve 0.86 2.23 NA NA 2.23 NA NA NA 1.6 NA
SDC20 4.44 4.52 10.59 4.53 10.59 11.53 4.66 4.36 NA 4.4 4.66
SDC22 1 1.6 1.56 1.26 1.46 1.56 1.5 0.93 1.56 1.36 1.6
SDC25 2 2.76 2.73 0.91 2 2.73 2.8 2.33 2.73 2.53 NA
SDC26 2.36 3.82 3.6 2.16 3.53 3.6 2.13 2.1 3.6 3.6 4.2
Grease
SDC12 15.9 14.5 23.7 19.6 22.4 23.7 22.3 15.9 21.8 14.5 21.9
SDC13 15.3 14.6 22.9 19.2 21.6 22.9 21.4 15.3 21.2 14.6 21.2
SDC41 5.16 13.8 19.93 2.3 2.3 19.93 11.63 4.86 20 12.7 11.63
SDC24 14.06 16.66 28.96 1 28.96 28.96 30.16 14 28.96 16.43 28.46
Note: the nuisance source tests identified with a ‘‘*’’ involve situations where notification to the building occupant is preferrable to remedy the situation and
prevent an ignition.
9
ARTICLE IN PRESS
10 L.A. Cestari et al. / Fire Safety Journal 40 (2005) 1–28
SDC07 SDC09
160 SDC10 SDC12
140 SDC13 SDC14
SDC33 SDC35
120 Nuisance Criteria SDC36 SDC38
100 If (dT/dt > 0.20) or (CO > 16) Then SDC39 SDC40
80 Al SDC20 SDC22
SDC25 SDC26
60
40
20
0
0.0 0.5 1.0 1.5 2.0
Rate of Temperature Rise, dT/dt (˚C/s)
180
SDC03
160 SDC04
140 SDC06
120 SDC08
Nuisance Criteria
SDC11
100 If (dT/dt > 0.20) or (CO > 16) Then Alarm
SDC31
80 SDC34
60 SDC37
40 SDC21
SDC23
20
SDC27
0
0.0 0.5 1.0 1.5 2.0
Rate of Temperature Rise, dT/dt (˚C/s)
Fig. 1. CO detector response versus rate of temperature rise for all tests. The algorithm that most closely
meets the design level of nuisance immunity is also shown.
ARTICLE IN PRESS
L.A. Cestari et al. / Fire Safety Journal 40 (2005) 1–28 11
The alarm times for Algorithms 1–3 are presented in Table 3 and Fig. 2.
Algorithm is at least as fire-sensitive as photoelectric detectors and is between
Algorithms 1 and 2. With the exception of one flaming fire (SDC 09), no significant
improvement in alarm time reduction is made from Algorithm 1. The cost of
reducing the detection time of SDC 09 is increasing the number of unwanted
nuisance alarms from zero-to-four.
Despite the delayed detection of 10 flaming fires (including all the four grease
fires), the nuisance immunity, sensitivity to smoldering fires, and fast detection of the
remaining flaming fires makes Algorithm 1 very promising. The addition of an
obscuration criterion to improve sensitivity to flaming fires is investigated in the next
section. The disadvantage of reducing detection times for the flaming fires using
Algorithms 2 and 3 is an increased level of nuisance sensitivity.
The response of the three sensors to each of the sources is provided in Fig. 3.
The nuisance algorithm is indicated by the rectangular solid and encloses virtually
all of the nuisance sources. The relative performance of this algorithm is presented
in Fig. 4.
This algorithm does not maintain the asymptotic nature of Gottuk et al. [8]. The
ionization threshold of k=0.15 m1 is based on the design criteria for nuisance
sensitivity. The CO threshold is inserted to maximize sensitivity to the smoldering
sources while still maintaining the design level of nuisance immunity. Also, it is
ARTICLE IN PRESS
12 L.A. Cestari et al. / Fire Safety Journal 40 (2005) 1–28
Fig. 3. Rate of temperature rise, CO detector response, and ionization detector output for all nuisance,
smoldering, and flaming tests.
ARTICLE IN PRESS
14 L.A. Cestari et al. / Fire Safety Journal 40 (2005) 1–28
Nuisance Criteria
(dT/dt> 0.20) or(CO>16) or (Ion >0.15)
Algorithm causes zero
Algorithm #4 Performance unwanted alarms
1.50
0.50
0.00
SDC01
SDC03
SDC04
SDC06
SDC08
SDC11
SDC31
SDC34
SDC37
SDC21
SDC23
SDC27
SDC02
SDC05
SDC07
SDC09
SDC10
SDC14
SDC33
SDC35
SDC36
SDC38
SDC39
SDC40
SDC20
SDC22
SDC25
SDC26
SDC12
SDC13
SDC41
SDC24
-0.50
-1.00
Ion Photo
0.50
0.00
SDC01
SDC03
SDC04
SDC06
SDC08
SDC11
SDC31
SDC34
SDC37
SDC21
SDC23
SDC27
SDC02
SDC05
SDC07
SDC09
SDC10
SDC14
SDC33
SDC35
SDC36
SDC38
SDC39
SDC40
SDC20
SDC22
SDC25
SDC26
SDC12
SDC13
SDC41
SDC24
-0.50
-1.00
Ion Photo
0.50
0.00
SDC01
SDC03
SDC04
SDC06
SDC08
SDC11
SDC31
SDC34
SDC37
SDC21
SDC23
SDC27
SDC02
SDC05
SDC07
SDC09
SDC10
SDC14
SDC33
SDC35
SDC36
SDC38
SDC39
SDC40
SDC20
SDC22
SDC25
SDC26
SDC12
SDC13
SDC41
SDC24
-0.50
-1.00
Ion Photo
A PCA is applied to the nuisance and fire test data. The previous analyses are
limited to developing algorithms utilizing a maximum of three or possibly four
sensors because visualizing the graphical representations of the raw data is difficult if
more than three dimensions are involved.
PCA is a technique that is applied to compress a large number of correlated
variables into a smaller number of uncorrelated variables while retaining most of the
‘‘important’’ information from the original set. So, PCA has the advantage of being
able to incorporate all eight available measurements into an algorithm by means of
data compression [10].
X ¼ t1 pT1 þ t2 pT2 þ . . . þ tk pTk þ E;
where, cov(X)pi=lipi
The process of conducting a PCA begins with calculating the loads from
an eigenvector decomposition of the covariance matrix calculated from ambient,
ARTICLE IN PRESS
L.A. Cestari et al. / Fire Safety Journal 40 (2005) 1–28 17
non-fire data.
2 3
pi;ion
6 7
6 pi;pho 7
6 7
6 7
6 pi;CO 7
6 7
6 7
6 pi;T 7
6 7
½ti ¼ ½X ion X pho X CO X T X d ion=dt X d pho=dt X dCO=dt X dT=dt 6 7:
6 pi;d ion=dt 7
6 7
6 7
6 pi;d pho=dt 7
6 7
6 7
6 pi;dCO=dt 7
4 5
pi;dT=dt
Mathematically, PCA decomposes the original data set X into the sum of products
of ti and pi plus a residual matrix. The ti’s, or ‘‘scores’’ are evaluated at each time
step, given the X’s representing the raw measurements, and pi’s, or ‘‘loads’’, for that
respective time step. The loads relate the measured variables to each other and the
scores relate the variables to the outcomes. This residual matrix contains any
remaining information from the original matrix that is not accounted for by the sum
of tipi products.
The ‘‘size’’ of the residual matrix E is a measure of how well the PCA model
predicts a given data matrix X. After the scores vectors are calculated at each time
step of the nuisance and fire test data, two different approaches are investigated for
fire detection algorithm development.
Application of the PCA results begins by monitoring the residual matrix E to
determine the point at which an atmosphere deviates from ambient conditions. An
abnormal situation is declared when the sum of the squares of each row, referred to
as the Q statistic of the residual matrix [11], exceeds a certain threshold (a Q value is
determined for each time step). Once an abnormal situation is declared the score
values tj at that time step are used to classify the deviation as being caused by a
nuisance or fire source. This approach is similar to that used by McAvoy, et al. [10]
who also investigated the efficacy of PCA for fire detection. Three consecutive data
points are required to declare an abnormal situation.
2.3.1. Algorithms accounting for scores throughout each test scores [t3/t4] algorithm
A detection algorithm using the t3 and t4 scores is strongly influenced by the
thermal and CO measurements as a result of the p3 and p4 loadings vectors [7]. The
algorithm for nuisance immunity is:
ARTICLE IN PRESS
Ion (std.) Photo (std.) PCA det Algorithm Algorithm Algorithm Algorithm Algorithm Algorithm Algorithm
time 10 nuisance 11 Ion— 12 nuisance 13 ion— 14 nuisance 15 ion 16 photo
(Q411.1) photo photo
Smoldering
SDC01 103.8 87.8 19.3 35.5 35.8 57.3 57.2 35.3 38 38.3
SDC03 NA NA 3.4 NA NA NA NA NA NA NA
SDC04 55.8 19.7 11.6 18.6 18.6 56 56 18.4 18.7 18.8
SDC06 100.1 57.4 28 56.7 56.7 100.2 100.2 56 56.9 57
SDC08 62.2 62.5 35.9 62.2 62.2 62.2 62.2 62.2 62.2 62.2
SDC11 71 70.9 22.9 62.7 62.7 70.9 70.9 59.9 64.8 65.5
SDC31 101.8 97.2 8.11 51.2 55.8 8.47 8.5 42.3 47.6 50.8
SDC34 65.6 NA 65 27.9 18.0 65.3 65.4 47.2 14.7 47.2
SDC37 26.2 23.8 28.4 25.9 26.9 3.17 28.4 24.4 26.4 25.9
SDC21 33.5 33.4 8.73 33.7 34.3 33.5 33.5 33.5 33.5 33.5
SDC23 74.2 25.9 26.8 28.2 29.5 35.1 34.2 28.2 28.5 28.5
SDC27 70.9 40.3 18.9 54.5 NA NA NA 68.8 67.7 67.7
Flaming
SDC02 1.4 2.1 1.1 2.7 1.4 2.2 1.5 2.2 1.5 1.4
SDC05 0.7 1.1 0.6 1.8 1 1.7 0.9 1.8 0.9 0.9
SDC07 1.3 No data 1.2 1.3 1.4 1.8 1.4 1.4 1.4 1.3
SDC09 3.1 3.6 2.9 3.1 3.1 3.5 3.1 3.2 3.1 3
SDC10 2.1 2.5 1.3 2.6 1.9 2.4 2.3 2.4 2.3 1.8
SDC14 57.5 57.4 57.1 57.9 57.2 57.8 57.5 57.9 57.5 57.5
SDC33 2.6 2.06 5.46 17.9 58.7 0.77 0.8 0.8 18.0 18.0
SDC35 1.08 2.03 1.56 0.13 1.57 1.5 0.57 0.57 0.53 0.17
SDC36 6.2 5.66 2.2 0.03 0.03 0.03 0.03 0.03 0.03 0.03
SDC38 11.12 4.9 2.08 0.03 0.03 0.03 0.03 0.03 0.03 0.03
SDC39 3.8 NA 1.96 1.5 1.53 1.7 1.57 1.6 1.6 0.07
SDC40 No data 0.86 4 0.03 0.03 0.03 0.03 0.03 0.03 0.03
SDC20 4.44 4.52 3.5 4.47 4.47 4.5 4.5 4.47 4.47 4.47
SDC22 1 1.6 0.6 1.13 1.9 1.37 1.13 1.27 1.1 1.03
ARTICLE IN PRESS
SDC12 15.9 14.5 3 21.8 21.8 23.4 19.6 21.7 16.7 15.4
SDC13 15.3 14.6 2.7 21.2 21.2 22.2 19.4 21.2 17.6 15.7
SDC41 5.16 13.8 4.6 2.97 11.6 11.6 4.9 2.93 4.9 4.77
SDC24 14.06 16.66 0.5 19.9 20.0 21.7 21.7 19.9 19.7 12.4
19
ARTICLE IN PRESS
20 L.A. Cestari et al. / Fire Safety Journal 40 (2005) 1–28
The performance of these two algorithms is indicated in Table 4 and Figs. 6 and 7.
The ellipse in Fig. 6 indicates the Algorithm 10 for nuisance immunity. Any point
outside of the ellipse relates to an alarm condition. As indicated in the top graph in
that figure, only those aggressive nuisance sources, where warning is deemed
important given the excessive obscuration, are outside of the ellipse.
The performance of these two algorithms is indicated in Table 4 and Figs. 8 and 9.
The [t2/t4] algorithm cannot meet the design criteria for both nuisance immunity and
fire sensitivity, perhaps due to the heavy weighting of the obscuration measurements
in the p2 loadings vector.
The [t3/t4/t2] scores algorithm required to satisfy the design criteria for nuisance
immunity and the algorithm required to most closely match the performance of a
conventional ionization and photoelectric detector are identified as Algorithms 15 and 16.
(15) Fire sensitivity—ionization detector: If 33.0[t3 cos(1.32)t4 sin (1.32)]2+1.8[t3
sin(1.32)+t4 cos (1.35)+32]2+0.9[t2+15]245000, then alarm
(16) Fire sensitivity—photoelectric detector: If 33.0[t3 cos(1.32)t4 sin (1.32)]2+1.8[t3
sin(1.32)+t4 cos (1.35)+32]2+1.3[t2]245000, then alarm.
ARTICLE IN PRESS
L.A. Cestari et al. / Fire Safety Journal 40 (2005) 1–28 21
Fig. 6. Scores t3 versus t4 for all nuisance, smoldering, and flaming tests. Algorithm that most closely
meets the design level of nuisance immunity is also shown.
The performance of these two algorithms is indicated in Table 4 and Fig. 10.
Nuisance Criteria
PCA Algorithm #10 Performance Algorithm causes zero
t3 - t4 unwanted alarms
1.50
Smoldering Flaming Grease
1.00
Relative Performance
0.50
0.00
SDCO4
SDC01
SDC03
SDC06
SDC08
SDC11
SDC31
SDC34
SDC37
SDC21
SDC23
SDC27
SDC02
SDC05
SDC07
SDC09
SDC10
SDC14
SDC33
SDC35
SDC36
SDC38
SDC39
SDC40
SDC20
SDC22
SDC25
SDC26
SDC12
SDC13
SDC41
SDC24
-0.50
-1.00
Ion Photo
0.50
0.00
SDCO4
SDC01
SDC03
SDC06
SDC08
SDC11
SDC31
SDC34
SDC37
SDC21
SDC23
SDC27
SDC02
SDC05
SDC07
SDC09
SDC10
SDC14
SDC33
SDC35
SDC36
SDC38
SDC39
SDC40
SDC20
SDC22
SDC25
SDC26
SDC12
SDC13
SDC41
SDC24
-0.50
-1.00
Ion Photo
algorithms. The averages are determined by calculating the average performance for
each set of sources (smoldering, flaming, and grease fires). In general, the algorithms
involving a CO threshold detected smoldering fires faster than conventional
detectors. The PCA algorithms that utilize t3, which is heavily weighted by CO,
also detect smoldering fires faster than conventional detectors. Also, the algorithms
involving the rate of temperature rise threshold provides detection times that are
similar or better relative performance than conventional detectors. Finally, all of the
algorithms investigated detected the grease fires slower than conventional detectors.
Most of the algorithms detected the grease fires shortly after flaming ignition.
Because frying food in oil is common in many residences, the transition of grease
sources from a nuisance source to a fire source is important. Algorithms for flaming
ARTICLE IN PRESS
L.A. Cestari et al. / Fire Safety Journal 40 (2005) 1–28 23
Fig. 8. Scores t2 versus t4 for all nuisance, smoldering, and flaming tests. Algorithm that most closely
meets the design level of nuisance immunity is also shown.
Nuisance Criteria
PCA Algorithm #12 Performance Algorithm causes zero
t2 - t4 unwanted alarms
1.50
0.50
0.00
SDCO4
SDC01
SDC03
SDC06
SDC08
SDC11
SDC31
SDC34
SDC37
SDC21
SDC23
SDC27
SDC02
SDC05
SDC07
SDC09
SDC10
SDC14
SDC33
SDC35
SDC38
SDC39
SDC40
SDC20
SDC22
SDC25
SDC26
SDC12
SDC13
SDC41
SDC24
SDC36
-0.50
-1.00
Ion Photo
0.50
0.00
SDCO4
SDC01
SDC03
SDC06
SDC08
SDC11
SDC31
SDC34
SDC37
SDC21
SDC23
SDC27
SDC02
SDC05
SDC07
SDC09
SDC10
SDC14
SDC33
SDC35
SDC38
SDC39
SDC40
SDC20
SDC22
SDC25
SDC26
SDC12
SDC13
SDC41
SDC24
SDC36
-0.50
-1.00
Ion Photo
One of the weaknesses of the currently available data set is the inability to assess
the repeatability and statistical variance of many of the sources tested. In particular,
data were available for only one test of each nuisance source. Understanding the
statistical variance of the fire sources is important for outside validation of the
current research and is also relevant to understanding the range of conditions that
occur during actual unwanted fires and nuisance scenarios. Another drawback of the
current data set is that the nuisance tests are limited to cooking sources, with the
exception of the smoking and tea candle tests. Examples of additional non-cooking
ARTICLE IN PRESS
L.A. Cestari et al. / Fire Safety Journal 40 (2005) 1–28 25
Fig. 10. Scores t3 versus t4 versus t2 for all nuisance, smoldering, and flaming tests. Algorithm that most
closely meets the design level of nuisance immunity is also shown.
ARTICLE IN PRESS
26 L.A. Cestari et al. / Fire Safety Journal 40 (2005) 1–28
nuisance sources could include shower steam, car exhaust, dust exposure (e.g. from
wood sanding), and cycling of an HVAC system.
3. Summary
The proposed algorithms are practical because they utilize relatively inexpensive
sensors that are commonly used in residential occupancies. Based on cost
considerations, a combined detector could be installed only in areas where nuisance
aerosols are expected, however, the advantages of early fire detection would then be
lost. The algorithm development procedures utilized here are general. Any of these
algorithms can be adjusted if the design criteria for nuisance immunity or fire
sensitivity are changed, i.e. a different balance is sought than applied in this paper.
Multi-sensor detection algorithms can provide higher levels of nuisance immunity
and fire sensitivity as compared to conventional ionization and photoelectric
detectors.
CO measurements detect smoldering fires earlier than conventional photoelectric
detectors. The CO detector also provides a much higher level of nuisance
immunity than either of the conventional smoke detectors.
The rate of temperature rise provides detection of flaming fires at least as fast as
conventional ionization detectors. The rate of temperature rise also provides a
greater level of nuisance immunity than either of the conventional smoke
detectors.
Both ionization and photoelectric detectors are found to be sensitive to nuisance
aerosols (i.e. prone to nuisance alarms). Increasing the ionization threshold
in a combined ionization/photoelectric detector is not effective in the current data
set.
The [rate of temperature rise/CO detector/ionization detector] algorithm provides
a high level of nuisance immunity and generally similar or earlier fire detection
compared to conventional detectors.
The current data set does not contain enough sensors to realize the advantages of
multivariate techniques, such as PCA.
Some loss in physical understanding becomes apparent in [t2/t4] algorithm
development.
The present study and numerous others have demonstrated the advantages that
advanced fire detection algorithms provide over conventional ionization and
photoelectric detectors. The use of additional sensors should be investigated. CO2
has shown promise in early detection of flaming fires [3,12]. Microsensor technology
can provide a large number of measurements on a relatively small chip (e.g. one
square inch). Fourier transform infrared (FTIR) spectroscopy can measure gaseous
species unique to individual smoke sources. The availability of additional source-
unique measurements has the potential benefit of increasing the ability of a multi-
sensor detector to properly classify signatures from fire and nuisance sources.
ARTICLE IN PRESS
28 L.A. Cestari et al. / Fire Safety Journal 40 (2005) 1–28
Acknowledgements
This work was a part of the Home Fire Smoke Detector Project and was funded
by the National Institute of Standards and Technology. The technical monitor for
the project was Richard Bukowski.
References
[1] Hall JR. The latest statistics on US home smoke detectors. Fire J 1989;83(1):39–41.
[2] Roby RJ, Gottuk DT, Beyler CL, Multi-signature Fire J US Patent Number 5,691,703, 1997.
[3] Milke JA. Monitoring multiple aspects of fire signatures of discriminating fire detection. Fire Technol
1999;35:195–209.
[4] Rose-Pehrsson SL, Hart SJ, Street TT, Tatem PA, Williams WW, Hammond MH, Gottuk DT,
Wright MT, Wong JT, Real-time probabilistic neural network performance and optimization for fire
detection and nuisance alarm rejection. Proceedings from the 12th international conference on
automatic fire detection ‘‘AUBE ‘01’’, March 2001.
[5] Gottuk DT, Peatross MJ, Roby RJ, Beyler CL. Advanced fire detection using multi-signature alarm
algorithms. Fire Safety J 2002;37:381–94.
[6] Bukowski RW, Peacock RD, Averill JD, Cleary TG, Bryner NP, Walton WD, Reneke PA,
Kuligowski ED. Performance of home smoke alarms: analysis of the response of several available
technologies in residential fire settings. Gaithersburg, MD: NIST; 2004.
[7] Cestari LA, Assessments of multi-components fire detection algorithms derived from nuisance and
real fire data. MS Thesis, Department of Fire Protection Engineering, University of Maryland, 2003.
[8] Gottuk DT, Wright MT, Wong JT, Pham HV, Rose-Pehrson SL, Hart S, Hammond M, Williams
FW, Tatem PA, Street TT, Prototype early warning fire detection system: test series 4 results. NRL/
MR/6180—02-8602, February, 2002.
[9] Bukowski RW, Field investigation of residential smoke detectors. Fire J 1997; 18.
[10] McAvoy TJ, Milke J, Kunt TA, using multivariate statistical methods to detect fires. Fire Technol.
1996; pp. 6–24.
[11] Wise BM, Gallagher NB. PLS_Toolbox 2.0 for use with MATLAB. Manson, WA: Eigenvector
Research, Inc.; 1998.
[12] Milke JA, McAvoy TJ. Analysis of signature patterns for discriminating fire detection with multiple
sensors. Fire Technol 1995;31(3):120–36.