Professional Documents
Culture Documents
Abstract
London Fire Brigade’s real fire library is a database of information collected from real fire
incidents by specialist teams of experienced fire investigators operating in the Greater London
Area. A sample of this data collected over a five-year period has been used to characterise the
distributions of fire sizes, fire growth rates and times between events that occur in building fires
in a form suitable for use with probabilistic risk assessment. The effect of occupancy type,
ignition source, first material ignited, and first aid fire-fighting by the occupants on the form of
these distributions was then examined. Incidents that produced very large losses, rapid growth
rates and extended time delays were also analysed to try to determine the reasons why such
extremes occur in real fires.
r 2004 Elsevier Ltd. All rights reserved.
Keywords: Fire investigation; Real fire data; Fire size; Fire growth rate; Log-normal distribution; Incident
database
1. Introduction
The cost of uncontrolled fires in buildings in terms of both human lives and
damage to property is a high one. Each year direct losses due to fires are estimated to
0379-7112/$ - see front matter r 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.firesaf.2004.05.002
ARTICLE IN PRESS
482 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
Nomenclature
account for between 0.1% and 0.4% of a country’s GDP [1,2]. To try to reduce such
losses it is important to be able to both understand and quantify the behaviour and
consequences of building fires in practice. The shift in emphasis from a prescriptive
to a performance based approach to fire safety engineering that has taken place in
recent years also means that a wide variety of data from real building fires is required
in order to be able to adequately quantify and assess the risk of fire (e.g. distributions
characterising fire damage, fire growth rates, times between events, etc.). Fire
investigation has the ability to provide such detailed information.
London is a city with a population of approximately seven million people.
It shares many similarities with other large cities around the world including
the problem of uncontrolled building fires. In London, teams of specialist fire
investigators are employed by London Fire Brigade working in shifts to provide 24 h
coverage of the 32 administrative areas (boroughs) of the Greater London Authority
together with the Corporation of the City of London (a total area covering
1578 km2). Their duties involve identification of the most probable causes and
sources of ignition of fires and the consideration of all issues involving fire
ARTICLE IN PRESS
P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524 483
development and the performance of fire protection. This is achieved through the
detailed examination of the fire scene, interviewing of witnesses and research into the
history of the individual buildings concerned.
Not all incidents would merit the attendance of a fire investigation unit (FIU),
since many fires are relatively minor and routine and do not warrant detailed
examination by specialist fire investigators. Thus, the investigators only attend fires
that meet certain criteria. A FIU automatically attends:
* Four pump fires and above, i.e. all fires where four or more fire engines are sent to
the scene of the fire.
* ‘‘Persons reported fires’’, i.e. fires where people are reported to be inside the
burning building when the call to the brigade is made.
The incident commander at the fire scene can also request the attendance of a FIU
at a fire based upon: human considerations (e.g. fatalities, serious injuries, rescues,
evacuations); operational considerations (e.g. fires where the cause or source of
ignition is suspicious or would otherwise be recorded as unknown); fire safety
considerations (e.g. buildings with automatic fire detection devices or fire
suppression equipment); fires of special interest and notifiable fires (e.g. new
materials and construction techniques).
Taken together such incidents represent around 25% of all the primary fires
attended by London Fire Brigade each year. These investigated fires effectively
represent the most ‘‘significant’’ incidents that have occurred.
Since 1996, the data collected by London Fire Brigade investigators at the
scene of each fire incident they attend has been entered into a database known
as the real fire library (RFL) [3,4]. The range of information collected into
the Library includes basic incident statistics (type of property, location, cause
of fire and source of ignition), details of the fire scene, fire development, fire
detection and protection and building egress. On average information from
investigated fires is collected into the library at a rate of around 4000 incidents per
annum.
In this paper a sample of data collected in the RFL in the five-year period between
1996 and 2000 has been analysed to try to characterise the distributions of fire-losses
(area of damage) and fire growth rates that occurred in both residential dwellings
and other types of buildings. The factors examined include:
* occupancy type
* ignition source
* effect of first aid fire-fighting by the occupants
* first material ignited.
The distributions of times between the ignition, discovery, call to brigade and
brigade arrival events have also been considered.
In addition, extreme cases (i.e. fires that produced very large losses, rapid growth
rates and extended time delays) were also examined to try to identify the reasons why
such extremes occur in real fires.
ARTICLE IN PRESS
484 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
2. Background
The heat release rate is one of the most important factors in the analysis of a fire
[17]. Examination of data from fire tests and real fires [18] suggests that the growth in
heat release rate during the early stages of a fire may often be reasonably
approximated by a time squared (t2 ) growth curve of the form
Q ¼ at2 ; ð1Þ
where Q is the heat release rate of the fire (kW), t the time after ignition (in s), and a
the fire growth parameter (kW/s2). Specific values of a have been selected to
characterise slow, medium, fast and ultra-fast fire growth rates [19,20].
ARTICLE IN PRESS
P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524 485
However, as Morgan [21,22] points out for a given occupancy there will actually
be a distribution of possible fire growth curves:
In reality, of course, any actual sample of fires occurring in the same nominal
occupancy will never be describable by a single growth curve. There will be a
distribution of growth curves depending on such factors as variations in fuel
layout and location of the initial ignitiony. In practice, however, the probability
distribution of growth curves is rarely known for an occupancy of interest to the
designer.
Ramachandran [23] has deduced such probability distributions for (exponential)
fire growth curves using UK fire statistics. Unfortunately the limited data available
meant that only a restricted range of occupancy types could be analysed. Morgan
notes that
ysimilar studies appear highly desirable for the major occupancies such as retail
and offices.
Working for the UK Home Office, Wright and Archer [24] analysed 1990s UK fire
data and derived relationships between the ‘‘age’’ of the fire (time from ignition until
fire brigade attendance at the fire) and the level of fire loss incurred (average area of
damage in m2) for a range of different building occupancy types. A linear regression
line was fitted to the data for each occupancy type, with the slope of the line
representing the rate of damage (m2/min) produced by a delay in attendance.
Occupancies were then categorised in accordance with this (linear) fire growth rate
* high: public buildings, factories and universities
* medium: retail, hotels and schools
* low: hospitals, licensed premises and offices
* very low: care homes.
Their review of previous research found that the application of fire growth models
to assigning fire cover was limited for a number of reasons, one of which was that
ythe distribution of fire sizes, rates of fire growth etc. has not been reported.
Wright and Archer also examined the level of fire loss incurred in terms of average
area of damage in m2 and financial loss (using insurance claim data) for a range of
different building occupancy types. Using this information they were able to derive a
relationship between fire brigade response time and financial fire loss for application
to fire cover risk assessment.
Sardqvist et al. [25] have previously examined a sample of the RFL data for fires in
non-residential premises in London 1994–1997. They obtained Complimentary
Cumulative Distribution Functions (CCDF) for the different time intervals and fire
areas recorded in the RFL and were able to use the data to derive correlations
between the water flow rate and total water applied to a fire and the fire damage
ARTICLE IN PRESS
486 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
area. The majority of fires they analysed did not spread after the arrival of the fire
brigade, while half did not grow any larger after they were discovered, due to self-
containment. They identified three categories of fire:
* Small self-contained fires that were confined to a single small object (1 m2 or less)
* Medium fires that spread in size by factor of up to ten after discovery but which
were contained upon fire brigade arrival (1–40 m2)
* Large fires that continued to spread in size after fire brigade arrival typically by a
further factor of ten (40–80 m2).
However, they did not consider the effect of different occupancy types on fire size.
3. Method
A sample of incidents containing suitable fire spread data was extracted from the
RFL for the five-year period between 1996 and 2000. The larger number and special
nature of fires in dwellings suggested that the fires in this occupancy type should be
analysed separately to those in other buildings. An overall sample of 2044 fires in
residential dwellings and 464 fires in other types of buildings was therefore used.
Data employed in the analysis included:
(i) Event times:
(a) time of ignition of the fire,
(b) time the fire was discovered,
(c) time the fire brigade was first called out,
(d) time of fire brigade arrival at the scene of the fire.
(ii) Fire areas:
(a) area of the fire at the time of discovery,
(b) fire area when the fire brigade arrived at the scene of the fire,
(c) final fire damaged area.
ARTICLE IN PRESS
P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524 487
However, not all of these items of data were recorded for every incident (i.e. some
items of data were missing for certain incidents).
An old computer science adage goes ‘‘garbage in–garbage out’’. This is equally
true for the data collected into the RFL. The quality of information obtained from
the RFL will only be as good as the data that has been entered into it and how well
that data is manipulated and retrieved once it is in the database.
The data entered into the database has effectively been filtered through specially
trained fire investigators. This should have provided a measure of quality control on
the data collected. However, by the very nature of the fire incidents involved, some of
the data collected, in some cases, will be subject to a degree of uncertainty and
subjectivity (this is particularly true of the time of ignition). Such uncertainties are
difficult to quantify, but where a sufficient quantity of data is available over an
ensemble of incidents, it is assumed here that statistically meaningful distributions
may still be obtained.
Inevitably, with the collection and entry of large amounts of data into the
database by many investigators, mistakes are made and errors occur. A validation
process has therefore been undertaken by the authors to try to identify and correct
those instances where the data is suspicious or incorrect.
Date and time data fields seem to be especially prone to error. It is all too easy to
transpose digits or day and month entries or forget to use the next day’s date when
an incident spans the times around midnight. Such date errors can render any
analysis using them meaningless, producing very large time differences, illogical
sequences of events and totally distort any statistics calculated using the erroneous
values.
While some of these date/time values are easy to identify (e.g. years like 1901)
most when taken in isolation appear perfectly valid. It is only when compared
with other date/times recorded for the same incident that the errors become apparent
(e.g. fire discovery and fire brigade arrival at dates spaced months apart). Routines
have therefore been implemented and used to check each of the date/time fields
stored in the relevant tables, relative to the other data/time fields in that table to
ensure that the values stored are self-consistent (i.e. check that the time intervals
between them are smaller than a specified threshold) and if appropriate that they
occur in a logical order (e.g. time of fire brigade arrival should not occur before time
of fire start).
Where possible cross-checks have also been made between fields in different tables
to check for consistency. Incidents where the data specified are inconsistent have
been identified and either corrected or removed.
ARTICLE IN PRESS
488 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
There are particular difficulties in estimating the actual time of ignition. The
specialist London Fire Brigade investigators rely on a combination of their
own judgement and the eyewitness interviews to provide the best possible estimate.
In many cases it was not possible to obtain a reliable estimate and such incidents
have therefore been excluded from analysis. However, in some cases it was
possible to give either a specific estimate or to set an upper and lower bound upon
the time of ignition. Such estimates represent the ‘‘best guess’’ of the ignition time
and should be treated with a degree of caution, particularly for fires which may have
smouldered for long periods and where the exact nature of the ignition ‘‘event’’ is
hard to define.
For dwellings where there was a relatively large sample of data available for a
single occupancy type only the incidents where a specific estimate of ignition time
was given by the investigators have been used in the analysis. However, in the case of
other buildings, where the total sample size available was smaller and divided over a
number of different occupancy types, a greater degree of uncertainty in the ignition
time has been allowed to increase the number of incidents open to analysis. Thus,
incidents where the difference between the upper and lower bound on the time of
ignition was reasonably small (15 min or less) were also included in the analysis. In
such cases the ignition time used in any calculations has been estimated as the
average time, midway between the upper and lower bounds specified.
As part of the analysis, other buildings were divided into the following occupancy
groups (similar to those used by Wright and Archer [24], but with the addition of
warehouses):
* care homes (for nursing the elderly and the mentally handicapped)
* factories (industrial premises, manufacturing)
* higher/further education (university, college, adult education)
* hospitals (general, psychiatric)
* licensed premises (e.g. public house, restaurants, bars etc.)
* offices
* public buildings (e.g. museums, art galleries, libraries)
* retail (shops, department stores, fast food restaurants)
* schools (infant, primary, secondary)
* warehouses (storage).
The (average) fire growth parameter, a (kW/s2) for each incident was estimated by
performing a least squares fit of a t2 growth curve based on areas of the fire when it
was discovered and when the fire brigade arrived and the time intervals between
ignition and discovery and ignition and fire brigade arrival, together with the
assumption that the fire area was zero at the time of ignition (i.e. the curve passes
ARTICLE IN PRESS
P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524 489
The distributions of fire sizes and fire growth parameter values observed were both
reasonably well approximated by the log-normal distribution.
Table 1
Fire growth parameter value classification scheme employed in the analysis
(a) (b)
(c) (d)
Fig. 1. Frequency histograms and cumulative probability plots for the fires investigated in residential
dwellings, comparing the distribution of the data sampled with the log-normal distribution fitted,
(a) frequency histogram of the natural log of fire damage area data along with normal distribution curve,
(b) probability plot of fire damage area data, (c) frequency histogram of natural log of fire growth
parameter data along with normal distribution curve, (d) probability plot of fire growth parameter data.
ARTICLE IN PRESS
P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524 491
deviation apparent at the tails. The deviation at small fire damage areas suggests that
the observed number of small fires is greater than would be expected according to the
log-normal distribution. Similarly, very large damage fires occur more frequently
than would be expected if they were distributed log-normally. It also apparent that
there are a series of ‘‘steps’’ in the distribution of data points. These steps reflect the
natural tendency of the investigators to round the fire damage area values recorded
(e.g. many fires with areas close to 1 m2 in size will be recorded as having an area of
1 m2 exactly).
The corresponding frequency histogram and cumulative probability plots found
for the fire growth parameter data (for the sample of fires investigated in residential
dwellings) are shown in Figs. 1(c) and (d). The log-normal distribution would also
appear to approximate the distribution of the fire growth parameter data reasonably
well. However, a measure of deviation is once again apparent at the tails of the
distribution, suggesting that very small fire growth rates occur more frequently and
very large fire growth rates occur less frequently than would be predicted if they were
strictly log-normal in form.
A similar pattern of (approximately) log-normal behaviour is found for the
distribution of both the fire damage area and fire growth parameter data, obtained
from the sample of fires investigated in other buildings, as shown in Figs. 2(a)–(d).
The histograms of the logarithm of both of the variables appear to be reasonably
normally distributed, whilst the data in both of the cumulative probability plots falls
close to the log-normal line. There is also generally less deviation apparent at the
distribution tails (particularly in the case of the fire growth parameter data) although
there is again evidence that very large damage fires (this time in other building types)
occur more frequently than would be otherwise predicted.
If x is log-normally distributed and z ¼ ln x, then z is normally distributed. The
log-normal distribution parameters (m; s) can therefore be estimated using
m ¼ meanðzÞ;
s ¼ standard deviationðzÞ:
The expected value of a log-normal distribution is the value that is most likely to
occur (i.e. the average fire size or growth rate) and is given by
s2
EðxÞ ¼ exp m þ ð3Þ
2
while the median value of the distribution is simply
x50 ¼ expðmÞ: ð4Þ
It is also useful to have some measure of the location of the tail of the distribution to
provide an indication of the extreme values that could occur. The 95th percentile of
the log-normal distribution provides a single parameter measure of the largest value
that will occur in 95% of cases and is given by [29]
x95 ¼ expðm þ 1:645sÞ: ð5Þ
ARTICLE IN PRESS
492 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
(a) (b)
(c) (d)
Fig. 2. Frequency histograms and cumulative probability plots for the fires investigated in other buildings,
comparing the distribution of the data sampled with the log-normal distribution fitted, (a) frequency
histogram of the natural log of fire damage area data along with normal distribution curve, (b) probability
plot of fire damage area datac (c) frequency histogram of natural log of fire growth parameter data along
with normal distribution curve, (d) probability plot of fire growth parameter data.
The complimentary cumulative distribution function (CCDF) for fire damage area
for the sample of fires investigated in residential dwellings is shown in Fig. 3 (the
CCDF gives the probability of a fire exceeding a particular fire size [29]). One-third
of the dwelling fires in the sample had a fire area less than or equal to 1 m2 in extent,
while the fire size was under 10 m2 in almost 80% of cases. However, there were also
some incidents in the tail of the distribution where a much larger fire damaged area
was achieved. In a number of these cases the fire propagated into a roof space or
ceiling void and was subsequently able to spread over a large area of the structure.
The distribution of fire sizes was found to be reasonably approximated by a log-
normal distribution (i.e. the natural log of fire size is approximately normal),
estimating the mean and standard deviation from the transformed natural log
variable (z ¼ ln x). The log-normal parameters characterising the distribution of fire
ARTICLE IN PRESS
P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524 493
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.01 0.1 1 10 100 1000 10000
Fire damage area (m2)
Fig. 3. The CCDF for fire damage area for the samples of fires investigated in both residential dwellings
and other buildings. Each distribution gives the probability of fire damage exceeding a given area.
sizes found for all dwelling fires in the sample are given in Table 2(a). The
corresponding expected fire size value found for this distribution using Eq. (3) was
7 m2, while the location of the 95th percentile of the distribution, x95 ; obtained with
Eq. (5) was 27 m2.
From the information collected by the fire investigators it is possible to
discriminate between those fires where the occupant took some form of first-aid
fire-fighting action against the fire (e.g. poured water over the fire or used a portable
fire extinguisher) and the fires where no such action was taken. The log-normal
distribution parameters, expected fire sizes and x95 values calculated for the two
corresponding dwelling fire size distributions are shown in Table 2(b). For the
distribution of dwelling fire damage areas where the occupants took no fire-fighting
action the expected fire size was 7 m2, but for the cases where the occupants did take
action the expected fire size was only 5 m2. The effect of occupant fire-fighting action
was also reflected in the x95 fire size values which was 27 m2 for the distribution
where no fire-fighting action was taken by the occupants, but 19 m2 for
the distribution where some form of action was taken by the occupants. A two-
sample t-test confirms that there is a statistically significant difference between the
means of the two log-normal distributions (po0:001), i.e. m is significantly lower for
fires where the occupant took some form of fire-fighting action.
It is also possible to discriminate between the dwelling fires investigated on the
basis of whether a smoke detector was installed in the property. The log-normal
parameters characterising these two types of distribution are shown in Table 2(c). A
two-sample t-test shows that the difference between the means of the two
distributions (and hence the median fire damage area) is statistically significant
(po0:001). However, both the expected fire damage and x95 damage values found
for the distribution of fire sizes where a detector was installed are only slightly lower
than those found for the distribution representing the cases where no detector was
fitted. This would suggest that installation of a smoke detector in dwellings does not
ARTICLE IN PRESS
494 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
Table 2
Log-normal parameters characterising the distribution of fire damage area in dwelling fires sampled
Group by
(a) All dwelling fires 1991 0.90 1.45 7 27
(a) All dwelling fires. (b) By whether fire-fighting action was taken by the occupant. (c) By whether a
smoke detector was installed. (d) By source of ignition.
The parameters m and s are the mean and standard deviation of the natural logarithm of the fire size.
The expected fire size of the distribution, EðxÞ; calculated using Eq. (3).
The fire size at the 95th percentile of the distribution, x95 ; calculated using Eq. (5).
a
Based on a sample of 1991 fires in other buildings where the final fire damage area was specified.
b
N is the number of fires in the group sampled.
in itself produce a large reduction in the amount of damage that can be typically
expected.
Intuitively, it might be expected that some forms of ignition source would be more
likely to give rise to larger fires than others. For example, fires due to domestic
appliances such as washing machines might often be contained within the framework
of the appliance and so not spread to other items. On the other hand fires started by
smoking materials left on an item of furniture might be expected to be more likely to
spread, resulting in larger fires.
This expectation does generally appear to be supported by the data as can be seen
by examining the expected fire-size and x95 values obtained for each of the ignition
source groups shown in Table 2(d). The fire size distribution for white goods (such as
washing machines, fridges and freezers) exhibits the smallest expected fire size (3 m2)
and x95 (12 m2) values. Perhaps more of a surprise is that the largest expected fire size
ARTICLE IN PRESS
P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524 495
(9 m2) and x95 (36 m2) values (apart from unknowns) are generated by the fire size
distribution found for incidents started by electrical supply and lighting equipment.
On the basis of the expected fire size and x95 values, the ignition sources can be
divided into two broad groups: those more likely to produce smaller fires (white
goods and cooking appliances) and those more likely to produce larger fires
(electrical supply and lighting equipment, naked flames, candles and smoking
materials). The former could be due to containment of the fire (by the case of the
appliance) or the likelihood of rapid discovery, while the latter group features
ignition sources, which would be more likely to spread to another item (different
forms of naked flame) or are likely to go for a long time before detection (electrical
supply and lighting, smoking materials).
Descriptive statistics for the time from ignition to discovery (and for the other time
intervals examined) in the dwelling fires sampled are summarised in Table 3.
The majority of the dwelling fires in the sample were discovered relatively quickly
with a median time from ignition to discovery of 4 min and 75% of the fires being
discovered in the 10-min period after ignition. However, there were also a number of
cases located in the tail of the distribution where there was a prolonged interval of
several hours between ignition and discovery of the fire. Such long discovery times
usually occurred in incidents where there was a smouldering fire (typically involving
smoking materials igniting settees, sofas or bedding) or where there was nobody in
the vicinity to discover the fire. For example,
* a smouldering towel covering hair tongs (an electrically operated hair styler) left
on by accident (7 h)
* a smouldering electric blanket (7 h)
Table 3
Descriptive statistics for selected time intervals in the dwelling fires sampled
Times interval (minutes) N Min Max Median Mean Std. dev. t95
a
Ignition and discovery 584 0 430 4 16.2 50.9 53
Discovery and call to FBb 1651 0 48 2 2.2 2.6 6
Call and arrival of FBc 1761 1 21 4 4.6 2.0 8
Ignition and FB arrivald 525 2 438 11 24.2 52.3 65
a
Based upon a sample of 584 fires in dwellings where specific ignition and discovery times were given.
b
Based upon a sample of 1651 fires in dwellings where the discovery and call times were specified and
the fire brigade took some form of action against the fire.
c
Based upon a sample of 1761 fires in dwellings where the call and arrival times of the fire brigade were
specified and the fire brigade took some form of action against the fire.
d
Based upon a sample of 525 fires in dwellings where the specific ignition and fire brigade arrival times
were supplied and the fire brigade took some form of action against the fire.
ARTICLE IN PRESS
496 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
For the dwelling fires sampled, the median time between discovery of the fire
and the call to the fire brigade was 2 min (Table 3). Almost half (44%) of the calls
to the fire brigade were made in the first minute after discovery of the fire while
95% of calls were made in under 6 min. There were also a small number of
incidents where there was a considerable delay between discovery of the fire and
the call to the fire brigade being made. In some of these cases a smouldering fire
Table 4
A breakdown of dwelling fires by source of ignition and time between ignition and discovery
Smoking materials 1 22 51 26 88
Other naked flames 25 42 29 4 171
Cooking appliances 23 40 35 2 107
Candles 4 43 48 6 54
Tv, hifi etc. 46 21 21 12 24
Electrical supply and 24 15 46 15 41
lighting
White goodsb 7 33 53 7 15
Heating appliances 30 19 40 11 57
Other sources 12 38 42 8 37
Unknown — 33 67 — 3
Allc 19 33 39 9 584
a
Rows may not add to 100% due to rounding error.
b
Includes washing machines, fridges, dishwashers, etc.
c
Total based on 584 fires investigated in dwellings where specific ignition and discovery times were
given.
ARTICLE IN PRESS
P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524 497
that was discovered and apparently extinguished by the occupant, was sub-
sequently found to be still burning and required the assistance of the fire brigade.
For example, in one incident a person placed a heater too close to the bed and
awoke to find the bed alight. They believed the fire to be extinguished, but woke
up nearly an hour later suffering with burns and the realisation that the fire
was still smouldering (thus the call to the brigade was made 48 min after the
fire was first discovered). In other cases a neighbour or passer-by discovered
the fire, but delayed calling the brigade either because they were not sure,
thought some one else would or because they did not know how to make such
a call.
Fig. 4 shows a comparison between the call time CCDF for cases where the
occupant took first-aid action against the fire with the CCDF distribution of call
times where no fire-fighting action by the occupant was taken. The distribution for
dwelling fires where the occupant took some form of fire-fighting action has a higher
proportion of fires exceeding a given call time. These results suggest that incidents
where the occupant took first-aid fire-fighting action were more likely to have a
longer delay between discovery of the fire and call to the brigade (i.e. evidence that
the use of first-aid fire-fighting action by the occupant can delay the call to the fire
brigade). A Man-Whitney U test [30] performed on these groups also implies that
there is a statistically significant difference in the location of these two distributions
(po0:001).
The median attendance time of the fire brigade to fires in dwellings was 4 min, with
95% of the incidents being attended within 8 min (Table 3).
0.9
0.8
No fire fighting by occupants
Probability of exceedance.
0.6
0.5
0.4
0.3
0.2
0.1
0
1 10 100
Time between discovery and call to brigade (mins)
Fig. 4. The CCDF comparing the time from discovery of fire to call to fire brigade being made for cases
where fire-fighting action was and was not taken by the occupants in residential dwellings.
ARTICLE IN PRESS
498 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
In the case of dwelling fires the median time between ignition and the arrival of the
fire brigade at the scene of the fire was 11 min (Table 3). There were also a number of
incidents where there was a significant length of time between ignition and
attendance of the brigade. These were primarily due to the long times between
ignition and discovery examined previously.
The frequency distribution of discovery times and call times were not readily
approximated by one of the standard distributions (e.g. normal, log-normal, etc.) for
which parameters could be estimated. As an alternative, the proportion of fires
exceeding a given time for each distribution is therefore tabulated in Table 5.
However, the distribution of Fire Brigade attendance times can be approximated
by a normal distribution using the mean and standard deviation parameters specified
(for time between call and arrival of the fire brigade) in Table 3.
The necessary data was available to allow t2 fire growth rate parameters to be
calculated (between ignition and arrival of the fire brigade) for 481 of the dwelling
Table 5
Proportion of fires in dwellings exceeding a specified time for selected time intervals
Time (min) Between ignition and Between discovery Between ignition and fire
discovery of firea of fire and call to brigade arrivalc
the fire brigadeb
fires sampled (using only those incidents where specific estimates of the ignition
and discovery times were given, the fire areas at discovery and brigade arrival
were specified and the fire brigade took some form of action against the fire).
Table 6 shows the growth rate parameter values of these fires classified in accordance
with the ranges specified in Table 1. Whilst the majority of dwelling fires had growth
rates that were either slow or very slow (smouldering), around 10% achieved
medium growth rates and 3% developed at rates that were fast (or in one case
ultra fast).
The frequency distribution of fire-growth parameter values, a; can also be
approximated using a log-normal distribution (i.e. such that ln(a) is normally
distributed). The log-normal distribution parameters estimated from the dwelling
fire-growth parameter data are shown in Table 7 along with the expected fire growth
parameter value, EðaÞ; (found using Eq. (3)) and the location of the 95th percentile,
a95 ; (found using Eq. (5)) for the distribution.
Table 7(a) gives the estimated log-normal distribution parameters (ma ; sa ) for the
dwelling fire growth parameter data sample treated as a whole. The expected fire
growth parameter value, EðaÞ; calculated for this distribution is 0.006 kW/s2 (just
below the threshold of classification for a medium growth rate fire) while the a95
value is 0.024 kW/s2.
The estimated distribution parameters, EðaÞ and a95 values for selected
ignition sources are shown in Table 7(b). Dwelling fires where the sources of
ignition were white goods or tv’s and hifi’s had the lowest expected growth rates and
a95 values, whilst fires started by heating appliances had the highest. The
distributions obtained for naked flames and candles had the highest ma values while
the deviation sa was highest for fires started by smoking materials and heating
appliances.
Table 7(c) presents the estimated distribution parameter values and EðaÞ values for
some selected types of first material ignited in dwelling fires. Dwelling fires where the
first material involved was a flammable liquid or vapour produced the highest
expected growth rate (0.023 kW/s2), a95 value (0.085 kW/s2) and ma value. The
deviation, sa ; was highest for the groups of fires where the materials first ignited were
upholstered furniture or rubbish.
Table 6
Number and percentage of dwelling fires sampled in each fire growth parameter class
Table 7
Log-normal parameters characterising the distribution of fire growth parameters in dwelling fires
Group by
(a) All dwelling fires 481 7.00 1.98 0.006 0.024
(a) All dwelling fires. (b) For selected sources of ignition. (c) For selected types of materials first ignited.
The parameters ma and sa are the mean and standard deviation of the natural logarithm of a:
The expected fire growth parameter of the distribution, EðaÞ; calculated using Eq. (3).
The value of the fire growth parameter at the 95th percentile of the distribution (Eq. (5)).
a
Based on a sample of 481 fires in dwellings where the value of a could be estimated, the fire brigade
took some form of action and a specific ignition time was estimated.
b
N is the number of fires in the group sampled.
Just over a quarter (26%) of the fires in all other buildings sampled (n ¼ 441;
where the final fire damage area was specified) had a fire damage area less than or
equal to 1 m2, while the fire size was less than 10 m2 in three-quarters (75%) of the
incidents investigated (see Table 8 and Fig. 3). However, there were also a number of
extreme cases where the fire damage was much larger, extending over hundreds or
even thousands of square metres.
Assuming that fire sizes can be modelled using a log-normal distribution, the
overall log-normal distribution parameters (m; s) estimated for fires investigated in
ARTICLE IN PRESS
P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524 501
Table 8
Percentage of fires in other buildings belonging to each fire damage size group by occupancy type
Care homes 65 29 6 — 17
Factories 13 55 23 9 47
Further education 29 43 29 — 14
Hospitals 47 50 3 — 30
Hotels 45 53 3 — 38
Licensed premises 26 56 16 2 50
Offices 30 44 22 3 63
Public buildings 12 53 24 12 34
Retail 17 53 27 3 94
Schools 35 38 26 — 34
Warehouses — 40 30 30 20
Allb 26 49 20 5 441
a
Rows may not add to 100% due to rounding error.
b
Total based on 441 fires investigated in other buildings where the fire damage area was specified.
all other buildings taken together are shown in Table 9(a), along with the expected
fire size (18 m2) and x95 value (66 m2).
As with dwelling fires, a comparison can be made between the distributions of fire
damage area in other buildings, based upon whether first-aid fire-fighting action was
taken by the occupant or not. The estimated log-normal parameters (m; s)
and expected fire size and x95 values found for the two distributions are shown in
Table 9(b). For the cases where the occupants took no fire fighting action, the
expected fire size was 24 m2, but for the cases where some form of action was taken
by the occupants the expected fire size was only 8 m2. A similar pattern was shown
by the x95 fire size values, which was 86 m2 for the distribution of fires where no
fire-fighting action was taken by the occupants, but just 30 m2 for the distribution of
fires where it was. A two-sample t-test confirms that the m value is significantly lower
for the cases where the occupant(s) took fire-fighting action against the fire
(po0:001).
A comparison can also be made between the fire size distributions found with and
without sprinklers (Table 9(c)). There was no significant difference in the means (and
hence the median fire size) found between the two distributions (a two-sample t-test
on the two means produces a p ¼ 0:861). However, both the expected fire size and
x95 values found were substantially lower for the distribution where a sprinkler
system was present, suggesting that sprinkler acts to curtail the occurrence of larger
fire sizes.
As with dwellings, it is once again possible to discriminate between the fires in
other buildings investigated on the basis of whether an Automatic Fire Detector
(AFD) system was installed or not. The log-normal parameters characterising these
ARTICLE IN PRESS
502 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
Table 9
Log-normal parameters characterising the distribution of fire damage area in other building fires sampled
Group by
(a) All other building fires 441 0.92 1.99 18 66
(a) All fires in other buildings. (b) By whether fire-fighting action was taken by the occupant. (c) By
whether sprinklers were installed. (d) By whether an AFD system was installed. (e) For some selected
sources of ignition.
The parameters m and s are the mean and standard deviation of the natural logarithm of the fire size.
The expected fire size of the distribution, EðxÞ; calculated using Eq. (3).
The fire size at the 95th percentile of the distribution, calculated using Eq. (5).
a
Based on a sample of 441 fires in other buildings where the final fire damage area was specified.
b
N is the number of fires in the group sampled.
two types of distribution are shown in Table 9(d). Both the expected fire size and x95
damage values found for the distribution of fire sizes in other buildings where an
AFD system was installed are less than a third of those found for the distribution
where no AFD system was present. A two-sample t-test also confirms that the large
difference between the means of the two distributions is statistically significant
(po0:001).
Table 9(e) gives the estimated log-normal parameters (m; s) together with the
expected fire sizes and x95 values found for some selected ignition sources starting
fires in other buildings. Based upon the expected fire size and x95 values, the ignition
sources can be broadly divided into three groups. The distributions obtained for
appliances (cooking, heating and other electrical) and other sources produced
ARTICLE IN PRESS
P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524 503
relatively small expected fire sizes of between 5 and 9 m2 (with x95 values of 35 m2 or
less). On the other hand the distributions found for fires started by smoking
materials and naked flames both had higher mean values, m; and consequently
produced larger expected fire sizes (18 and 21 m2, respectively) and x95 values (67 and
78 m2). However, by far the largest expected fire size, at 85 m2, and x95 value (195 m2)
was found for fires started by electrical supply and lighting equipment (as was also
the case in dwellings). This was primarily as a consequence of a large standard
deviation s; suggesting that this type of ignition source was responsible for a wide
range of fire sizes—some small, others extremely large.
A breakdown of fire size by occupancy group is also given in Table 8. It is evident
that the largest fire sizes occurred in warehouses, with 30% (6 fires, n ¼ 20) of the
fires investigated in this occupancy group damaging areas in excess of 100 m2. There
were also a significant number of very large fires (greater than 100 m2) that occurred
in factories and public buildings. In contrast the majority of the fires that were
investigated in care homes, hospitals and hotels were less than 1 m2 with few
exceeding 10 m2 in area.
The log-normal distribution parameters (m; s) estimated for fires occurring in
each type of occupancy are shown in Table 10 along with the expected fire size and
95th percentile (x95 ) values calculated using Eqs. (3) and (5). The expected fire size
found for warehouses, at 170 m2, is significantly higher than that found for any of the
other occupancies, as is the x95 value (586 m2) and the mean m: In contrast, the
expected fire sizes, x95 and m values found for care homes, hospitals and hotels are all
significantly lower than those found for the other occupancy groups.
Table 10
Log-normal parameters characterising the distribution of fire damage area for different occupancy groups
Occupancy group
Care homes 17 0.64 1.44 1 6
Hospitals 30 0.08 1.28 2 8
Hotels 38 0.29 1.52 2 9
Licensed premises 50 0.78 1.70 9 36
Schools 34 0.69 1.89 12 45
Retail 94 1.17 1.84 18 66
Higher/further education 14 0.56 2.24 22 70
Offices 63 0.83 2.14 23 78
Factories 47 1.68 1.91 34 124
Public Buildings 34 1.80 1.92 38 142
Warehouse 20 2.87 2.13 170 586
The parameters m and s are the mean and standard deviation of the natural logarithm of the fire size.
The expected fire size of the distribution, EðxÞ; calculated using Eq. (3).
The fire size at the 95th percentile of the distribution, calculated using Eq. (5).
a
Based on a sample of 441 fires in other buildings where the final fire damage area was specified.
b
N is the number of fires in the group sampled.
ARTICLE IN PRESS
504 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
To increase the number of usable incidents for other buildings, some cases in
which there was a greater degree of uncertainty in the ignition time were included
using the average time between the upper and lower bounds on the ignition time
specified for incidents where the difference between the two bounds was fifteen
minutes or less.
Descriptive statistics for the time from ignition to discovery (together with the
other time intervals examined) for fires in other buildings sampled are summarised in
Table 11.
The median time from ignition to discovery of fires in other buildings was 4 min
(same as for dwellings) with nearly 80% of the fires being discovered in 10 min or
less. However, there were also a small number of incidents in the tail of the
distribution where the fire took more than 30 min to be discovered. Reasons for such
long discovery times identified included smouldering cigarettes being left on a sofa or
bed or the building being left unoccupied overnight.
A breakdown of discovery time by occupancy group is shown in Table 12.
A relatively high percentage of the fires in factories were discovered at ignition
(35%) typically because they involved some form of explosion that would be
detected immediately. Fires in retail premises also tended to be more frequently
discovered at ignition time (28%) often because they involved an accident with
cooking appliances that were being used by the occupier at the time of the fire.
In contrast fires in office buildings had a higher incidence of discovery times in
excess of 30 min (16%) than the other occupancies, largely as a result of cases where
there was a smouldering fire in an office started by cigarettes or other smoking
materials.
Table 11
Descriptive statistics for selected time intervals in the other building fires sampled
Times interval (minutes) N Min Max Median Mean Std. dev. t95
a
Ignition and discovery 215 0 599 4 10.3 42.2 30
Discovery and call to FBb 342 0 24 2 1.9 2.1 5
Call and arrival of FBc 369 1 13 4 4.4 1.6 7
Ignition and FB arrivald 180 3 606 10 17.8 46.3 40
a
Based upon a sample of 215 fires in other buildings where the ignition and discovery times were
specified and the ignition time range estimated was less than or equal to 15 min.
b
Based upon a sample of 342 fires in other buildings where the discovery and call times were specified
and the fire brigade took some form of action against the fire.
c
Based upon a sample of 369 fires in other buildings where the call and arrival times of the fire brigade
were specified and the fire brigade took some form of action against the fire.
d
Based upon a sample of 180 fires in other buildings where the ignition and fire brigade arrival times
were specified, the ignition time range estimated was less than or equal to 15 min and the fire brigade took
some form of action against the fire.
ARTICLE IN PRESS
P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524 505
Table 12
A breakdown of discovery time by occupancy group for fires in other buildings
Care homes 6 53 41 — 17
Factories 35 20 35 10 20
Higher/further 12 50 38 — 8
education
Hospitals 5 62 33 — 21
Hotels 17 56 28 — 18
Licensed Premises 14 36 45 5 22
Offices 4 44 36 16 25
Public buildings 9 18 64 9 11
Retail 28 26 41 4 46
Schools — 40 55 5 20
Warehouses — 43 57 — 7
Allb 14 39 41 5 215
a
Rows may not add to 100% due to rounding error.
b
Total based on 215 fires investigated in other buildings where the ignition and discovery times were
specified and the ignition time range estimated was less than or equal to 15 min.
Table 13
A breakdown of other building fires by source of ignition and time between ignition and discovery
Cooking appliances 33 36 27 3 33
Heating appliances 18 18 65 — 17
Other appliances 8 58 35 — 26
Smoking materials — 27 45 27 22
Naked flames 10 46 42 2 89
Electrical 24 29 38 10 21
distribution
Other sources 25 25 50 — 4
Unknown — — 100 — 3
Allb 14 39 41 5 215
a
Rows may not add to 100% due to rounding error.
b
Total based on 215 fires investigated in other buildings where the ignition and discovery times were
specified and the ignition time range estimated was less than or equal to 15 min.
The influence of the ignition source upon the time to discovery of fires in other
buildings is summarised in Table 13. While a third of the fires involving cooking
appliances were discovered at ignition, only 3% of these fires took more than 30 min
to be discovered. Conversely, none of the fires started by smoking materials were
ARTICLE IN PRESS
506 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
discovered at ignition, but 27% of these fires took more than 30 min to be
discovered.
Table 11 also shows the descriptive statistics found for the time interval between
the discovery of the fire and the call being made to alert the fire brigade in other
building fires (note that only those incidents where the fire brigade attended and took
some form of fire-fighting action against the fire were considered). The median time
between discovering a fire in other types of building and calling the Fire Brigade was
2 min (the same as in dwellings). Almost half of the calls to the fire brigade were
made in the first minutes, while 95% of calls were made in the first 5 min. However,
there were also a small number of incidents where there was a significant delay, in
excess of 5 min, between discovering a fire and calling the brigade. Reasons for a
delay in calling the fire brigade identified were:
The median attendance time of the fire brigade to fires in other buildings was
4 min, with 95% of the incidents being attended within 7 min (Table 11).
5.5. Time from ignition to arrival of the Fire Brigade (fire age)
For fires in other buildings the median time between ignition and arrival of the fire
brigade was 10 min, while 95% of the intervals between ignition and brigade arrival
did not exceed 40 min (Table 11).
0.9
0.6
0.5
0.4
0.3
0.2
0.1
0
1 10 100
Time between discovery and call to brigade (mins)
Fig. 5. The CCDF comparing the call times for cases where fire-fighting action was and was not taken by
the occupants in other buildings.
ARTICLE IN PRESS
508 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
The frequency distribution of discovery times and call times were not readily
approximated by one of the standard distributions (e.g. normal, log-normal, etc.) for
which parameters could be estimated. As an alternative, the proportion of fires
exceeding a given time for each distribution is therefore tabulated in Table 14.
However, as with dwelling fires, the distribution of Fire Brigade attendance times
can be approximated by a normal distribution using the mean and standard
deviation parameters, specified in Table 11, for the time interval between the call and
arrival of the fire brigade.
The requisite data was available to allow t2 fire growth parameter values to be
calculated (between ignition and arrival of the fire brigade) for 164 of the fires
investigated in other buildings. Table 15 shows the growth parameters values of these
fires classified in accordance with the ranges specified in Table 1, together with a
breakdown by occupancy group. Whilst the majority (81%) of the fires in other
buildings were either slow or very slow (smouldering), 10% developed at rates
that could be classed as being fast or ultra fast. Such fast and ultra-fast growth
fires occurred mainly in warehouses (3, n ¼ 6), factories (4, n ¼ 16) and retail
Table 14
Proportion of fires in other buildings exceeding a specified time for selected time intervals
Time (min) Between ignition and Between discovery of fire Between ignition and fire
discovery of firea and call to the fire brigadeb brigade arrivalc
Table 15
Number of fires in other buildings belonging to each fire growth parameter class by occupancy group
Occupancy group Very slow Slow Medium Fast Ultra fast Total
Care homes 5 4 — — — 9
Factories 4 7 1 3 1 16
Higher/further education 1 4 — — — 5
Hospitals 7 10 — — — 17
Hotels 5 7 — — — 12
Licensed premises 4 10 2 — 1 17
Offices 6 11 2 — — 19
Public buildings 2 6 1 1 — 10
Retail 4 18 9 4 2 37
Schools 6 9 — 1 — 16
Warehouse — 2 1 2 1 6
Other building fires (all) 44 88 16 11 5 164
Other building fires (%) 27% 54% 10% 7% 3% 100%
a
Based on a sample of 164 fires in other buildings where the value of a could be estimated, the fire brigade
took some form of action and the uncertainty in the time of ignition was 15 minutes or less.
premises (6, n ¼ 37). In contrast, only slow or very slow growth rate fires were
recorded in care homes, hospitals, hotels and further education premises (although
the latter suffers from a very limited sample size).
As with dwellings, the frequency distributions of fire growth parameter values, a,
were once again modelled using a log-normal distribution. The overall log-normal
distribution parameters (ma ; sa ) estimated for fires investigated in all other buildings
taken together are shown in Table 16(a), along with the expected fire growth
parameter value (0.012 kW/s2) and a95 value (0.044 kW/s2) equivalent to a medium
growth rate fire.
Table 16(b) gives the estimated log-normal parameters (ma ; sa ) together with the
expected fire growth parameter and 95th percentile (a95 ) values found for some
selected types of ignition source starting fires in other buildings. Fires that were
started by electrical supply and lighting equipment exhibited the highest expected
growth rate (0.043 kW/s2) and a95 (0.134 kW/s2) with the distribution displaying both
the greatest mean (ma ) and standard deviation (sa ) values. The expected fire growth
rate and a95 values obtained for fires started by cooking appliances, at 0.029 and
0.096 kW/s2, respectively, were also relatively high.
The estimated log-normal distribution parameters and expected fire growth and
a95 values for selected types of first materials ignited in fires in other buildings are
shown in Table 16(c). As was the case in dwellings, fires where the first material
involved was a flammable liquid or vapour produced the highest expected growth
rate (0.070 kW/s2), a95 value (0.231 kW/s2) and mean (ma ) value.
The log-normal distribution parameters (ma ; sa ) estimated for fires occurring in
each type of occupancy are shown in Table 17 along with the expected fire size and
a95 : The expected fire growth parameter value found for warehouses, at 0.107 kW/s2,
ARTICLE IN PRESS
510 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
Table 16
Log-normal parameters characterising the distribution of fire growth parameters in the other building fires
sampled
Group by
(a) All other building fires 164 6.5 2.0 0.012 0.044
(a) All other building fires. (b) For selected sources of ignition. (c) For selected types of materials first
ignited.
The parameters ma and sa are the mean and standard deviation of the natural logarithm of a:
The expected fire growth parameter of the distribution, EðxÞ; calculated using Eq. (3).
The value of the fire growth parameter at the 95th percentile of the distribution (Eq. (5)).
a
Based on a sample of 164 fires in other buildings where the value of a could be estimated, the fire
brigade took some form of action and the uncertainty in the time of ignition was 15 min or less.
b
N is the number of fires in the group sampled.
is significantly higher than that found for any of the other occupancies, as are the a95
mean ma : Conversely, the expected fire growth parameter, a95 and ma values found for
care homes, hospitals and higher and further education are all lower than those
found for the other occupancy groups.
6. Discussion
There were a relatively small numbers of fires in dwellings that resulted in large fire
damage areas, particularly in cases where the fire had been able to spread into a roof
space or ceiling void. In other buildings the largest fire damage areas occurred in
warehouses which by their very nature often present a large open space through
which fire can spread with relative ease. From an examination of the incidents
ARTICLE IN PRESS
P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524 511
Table 17
Log-normal parameters characterising the distribution of fire growth parameters for each occupancy
group in the other building fires sampled
Occupancy group
Care homes 9 7.7 1.1 0.001 0.003
Higher/further education 5 7.2 0.8 0.001 0.003
Hospitals 17 7.1 1.3 0.002 0.007
Hotels 12 7.7 2.1 0.004 0.014
Offices 19 7.1 1.8 0.004 0.016
Schools 16 7.3 2.0 0.005 0.019
Public buildings 10 6.2 1.9 0.012 0.045
Licensed premises 17 6.6 2.2 0.016 0.053
Retail 37 5.4 1.9 0.027 0.101
Factories 16 5.9 2.2 0.030 0.100
Warehouse 6 4.0 1.9 0.107 0.405
The parameters ma and sa are the mean and standard deviation of the natural logarithm of a:
The expected fire growth parameter of the distribution, EðaÞ; calculated using Eq. (3).
The value of the fire growth parameter at the 95th percentile of the distribution (Eq. (5)).
a
Based on a sample of 164 fires in other buildings where the value of a could be estimated, the fire
brigade took some form of action and the ignition time range estimated was 15 min or less.
b
N is the number of fires in the group sampled.
involved, reasons identified for the occurrence of large fire sizes in other types of
building included:
* Fires involving sandwich panel cladding, promoting rapid flame spread and
making it difficult to detect and fight the fire (warehouses);
* Fires in premises engaged in manufacturing and storage of fuel sources with a fast
heat release rate such as plastic furniture or other plastic products (warehouses
and factories);
* Explosions triggered by the ignition of flammable vapours released by solvents
(factories);
* Involvement of highly combustible packaging material such as polystyrene
(warehouse and retail);
* Ignition of PU foam settees (hotels and public buildings);
* Fire spread into a timber roof space where it was able to extend over a large area
(licensed premises, offices and schools);
* Use of petrol or other accelerants in deliberate fires (licensed premises, retail);
* Inadequate fire separation between compartments allowing fire spread (all).
One perhaps surprising result of the analysis, was the large fire damage areas
(indicated by the expected fire size and x95 values) found for distributions of fires
ARTICLE IN PRESS
512 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
started by electrical supply and lighting, relative to the other sources of ignition,
particularly naked flames (other) where some form of deliberate action was usually
involved.
In dwellings, this would appear to be primarily due to electrical faults (e.g. wiring
overheating, problems with power supplies, etc.) causing fires in roof spaces. Such
roof voids often provide a large undivided space through which a fire can spread
relatively easily over a significant area, while the timber roof construction typically
used, offers an extensive fuel source to facilitate fire growth. It would also seem
likely that fires started by electrical supply and lighting in a roof space (i.e. in a
relatively remote part of the dwelling above the siting of many domestic smoke
alarms) would generally take longer to be discovered and hence have the opportunity
to cause a larger amount of damage.
The situation is less clear-cut in the case of other buildings, since the high expected
fire size and x95 values obtained for fires started by electrical supply and lighting are
based on only a relatively small number of incidents that caused a large amount of
damage in the tail of the distribution. Thus, it is difficult to determine if the result is
genuine or just a consequence of the limited sample size and should be treated with a
degree of caution.
To illustrate this, Fig. 6 shows a comparison of the boxplots characterising the
distribution of fire sizes found for fires in other buildings started by electrical supply
and lighting and other naked flames. While the boxplot obtained for other naked
flames shows a high concentration of incidents at moderately large fire sizes, the
boxplot found for electrical supply and lighting reveals a much smaller number of
cases extending over a wider range of fire sizes. Thus while the distribution found for
1200
1000
Fire damage area (square metres)
800
600
400
200
0
N= 52 163
Electrical supply Other naked flames
Fig. 6. Comparison of boxplots showing the distribution of fire damage area obtained for fires started by
other naked flames with that found for fires started by electrical supply and lighting sources. The asterisks
denote the outliers of each distribution located at large area values.
ARTICLE IN PRESS
P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524 513
other naked flames has the higher mean value, that obtained for electrical supply and
lighting has a more extended tail.
One possible explanation for the results found could be that electrical supply and
lighting is a popular scapegoat for the cause of large fires started in other buildings
where there has been extensive damage and the source of ignition is more difficult to
determine. However, Hall [31] has observed that there is no evidence to support such
a bias occurring in practice:
If the effect is a genuine one, then on the basis of an examination of the limited
number of incidents sampled it is possible to speculate that the large fire sizes in the
tail of the distribution may at least in part be attributed to fires started by electrical
wiring or light fittings in services and concealed spaces (e.g. lift shaft, sandwich
panels, etc.). Such fires, embedded within the structure of a building have the
potential to spread over large areas relatively easily and would be more difficult to
both detect and fight.
Much less surprising were the high fire growth rates attributable to the ignition of
flammable liquids and gases found for both dwellings and other buildings. Such
rapid growth rates reflect the rapid nature of the combustion process in flammable
vapours. The results also support the observation made in DD240 [20] that
Most fires that do not involve flammable liquids or gases initially grow relatively
slowly.
In both dwellings and other buildings the expected fire size and x95 values were
smaller for the distribution of fires where the occupants took some form of first-aid
fire-fighting action against the fire. These results might suggest that occupant fire-
fighting actions can reduce the level of fire damage incurred. However, it is possible
that they simply reflect a tendency for building occupants to only attempt to tackle
relatively small fires.
It is also possible to examine the effect of the fire size at discovery of the fire on the
success of first-aid fire-fighting action by occupants in extinguishing fires by
considering whether the fire brigade subsequently took further action against the fire
or not. Thus, in cases where no further action was taken by the brigade, the first-aid
ARTICLE IN PRESS
514 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
measures are deemed to have been a success while in cases where further action by
the brigade did occur they are assumed to have failed.
Table 18 shows the log-normal parameters characterising the distributions of fire
area at discovery on the basis of success (or failure) of first-aid fire-fighting measures
for both dwellings and other buildings. In dwellings, both the expected and x95
values were lower for fires where first-aid fire-fighting action was a success. A two-
sample t-test confirms that the m value was significantly lower for the distribution
where the occupant’s fire-fighting action against the fire was successful (po0:01).
This pattern was repeated for fires in other buildings.
On the basis of the x95 results it can be observed that the majority of fires in
dwellings that were successfully extinguished by the occupants were under 3.1 m2 in
area when they were discovered. Similarly, in other buildings, the majority of fires
that were successfully extinguished by the occupants were less than 2.3 m2 in area or
when they were discovered. This would suggest that fires that are of greater size when
they are discovered are unlikely to be successfully extinguished by occupant fire-
fighting action alone.
The results suggest that the installation of smoke detectors in dwellings only
produced a relatively modest reduction in the amount of fire damage that could be
typically expected. This can be attributed to the large number of detectors installed
in dwellings that did not actuate when a fire occurred (approximately 50% in the
sample examined here) due primarily to a battery being flat or having been removed
altogether. To illustrate the consequences of this, Table 19 shows the log-normal
parameters characterising the distribution of fire sizes for dwellings where a detector
Table 18
Log-normal parameters characterising the distributions of fire area at time of discovery on the basis of
success (or failure) of first-aid fire-fighting measures for both dwelling and other building fires
Dwelling firesa
Fire fighting action by occupant
Success 92 0.69 1.11 0.9 3.1
Failure 240 0.29 1.26 1.7 6.0
Table 19
Log-normal parameters characterising the distributions of fire damage area for fires in dwellings where a
smoke detector was installed on the basis of whether the system was actuated or not
was installed and the detector either did or did not actuate. The distribution
representing the cases where the detector did actuate has expected fire size and x95
damage values that are substantially lower than (less than half) those found for the
cases where the detector did not actuate. Of course the primary purpose of smoke
detectors is life-saving, not property loss reduction. Thus, while any reduction in fire
damage would be welcome, detectors should not be derogated on this basis, even if
no significant reduction is apparent.
Based upon the log-normal mean parameter values and expected fire sizes found
for each distribution the occupancies can be divided into four broad fire damage
(loss) classes:
These results emphasise the very high fire loss values that occurred in warehouse
(storage) premises.
Wright and Archer [24] give figures (obtained from UK fire data) for the average
fire damage areas that occurred in a number of different occupancies (note that these
simple arithmetic average area values may be skewed to the right due to the nature of
the distribution). These results are largely consistent with the fire damage classes
identified above (e.g. factories and public buildings both had relatively high average
damage areas whilst those found for care homes and hospitals were relatively low).
The one real exception to this is the average damage area found for schools which
has a relatively high value (similar to that found for public buildings) inconsistent
with being classified in the medium fire loss class.
Unfortunately, the limited amount of data available to characterise the fire growth
parameter distributions for some of the occupancies, means that only a broad
ARTICLE IN PRESS
516 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
classification is really justified. Based upon the log-normal mean values found for
each distribution the occupancies can be divided into two fire growth classes:
* high: warehouses, factories, public buildings and retail;
* low or medium: care homes, further education and hospitals, hotels, offices,
schools, licensed premises and dwellings.
These relative classifications are generally consistent with those made by Wright
and Archer (based upon area growth rates derived using linear regression), where
fires in public buildings and factories were also identified as likely to produce high
growth rates, while those occurring in care homes and hospitals were comparatively
low. However, there are also some differences. While fires in universities were
identified as producing a high rate of damage by Wright and Archer, the results
here suggests that the growth rate of fires in higher/further education establishments
was (on average) relatively low. This discrepancy may be attributable to the very
limited sample of incidents available for this occupancy type (only 5 fires) suggesting
the values estimated to characterise the distribution in this case might be unreliable.
The results presented here also suggest that retail should be classified as producing
high growth rate fires, while the study made by Wright and Archer identifies this
occupancy as being medium. At least some of this variance may be due to the
difference in the way fire growth rate was characterised in terms of linear area
growth rate by Wright and Archer, against t2 heat release rate here and the use here
of a higher heat release rate (per unit area) for this occupancy type.
The circumstances surrounding the fires investigated in dwellings with the highest
fire growth rates are summarised in Table 20. Using this information it is possible to
identify a number of scenarios in which fires with high growth rates occurred.
6.7.4. Candles left burning unattended and too close to other items
Incidents where items were left in close proximity to a candle and which became
involved.
ARTICLE IN PRESS
P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524 517
Table 20
Circumstances surrounding the fires with the highest fire growth parameter (a) values investigated in
dwellings
0.130 0.8 Deliberate suicide attempt by an elderly person igniting mains gas with the
piezo electric starter on a cooker.
0.074 — A workman using a blow lamp in the roof space of a house unintentionally
ignited timber soffit boards. The resulting fire spread rapidly throughout the
roof space.
0.069 1 Unintentional disposal of a cigarette on a double divan mattress.
0.053 0.6 Unintentional disposal of a cigarette on a Polyurethane foam sofa.
0.042 2 Gas leak ignited by a naked flame after a failed attempt to fit a cooker.
0.039 — Butane gas leak from a domestic blow lamp in an utility room.
0.039 1 Occupier unintentionally ignited excess paraffin whilst lighting a portable
heater in the living room.
0.035 1 Rubbish in bedroom unintentionally ignited whilst using paint-stripper. Fire
subsequently spread into roof space.
0.034 0.9 Candle left alight whilst occupier fell asleep in bed-sitting room, ignited the
duvet cover. The Fire then spread rapidly to cushions and curtains.
0.031 — Fire started when child playing with matches in the living room ignited
curtains and rapidly spread to involve room contents.
0.031 1 The occupier was trying to thin gloss paint by warming it on a gas cooker
hob. The flammable vapour released collected beneath the ceiling and
triggered an explosion when the layer spread down to the level of the gas
ring.
0.030 4 Fire started by a child playing with matches ignited tissue paper in the living
room and rapidly spread to settee.
0.030 — A radiant heater ignited an armchair that was placed too close. The fire
spread rapidly to involve the room furniture.
0.029 2 Cooking oil ignited in a chip pan left unattended in the kitchen.
0.023 — Spillage of paraffin from lighted oil heater in dining room was
unintentionally ignited and rapidly involved the room contents.
0.023 2 Clothes left to dry in the living room were placed too close to a radiant
heater and were ignited. The fire then spread rapidly to involve the room
furniture.
0.022 0.8 The occupier was burning off paint with a blow lamp in the bedroom and
unintentionally ignited the timber window frame. The resulting fire rapidly
involved the room contents.
0.022 1 A youth under the influence of alcohol ignited some paper whilst playing
with a cigarette lighter in the bedroom. The fire then involved the carpet and
spread rapidly.
0.021 2 A child playing with a naked flame ignited some boxes under the staircase.
0.021 1 A candle in the living room ignited a book that had been left too close and
rapidly spread to involve the room contents.
The estimated fire size doubling time, tdb ; for each incident is also given (where available).
The table also shows the estimated fire size doubling time, tdb ; (i.e. the time
required for the fire to double in area) for incidents where the necessary data was
available, based upon the growth (in fire area) between discovery and arrival of the
fire brigade. If the growth in fire area is assumed to be exponential in form, then the
fire size doubling time can be estimated using
Dt12 lnð2Þ
tdb ¼ ; ð6Þ
lnðA2 =A1 Þ
where A1 is the area of the fire when it was first discovered (m2); A2 the area of the
fire when the fire brigade arrived (m2); and Dt12 the time interval between discovery
of the fire and arrival of the fire brigade (min).
Wright and Archer [24] suggest that for uncontrolled fires, a fire with a doubling
time of 4 min can be equated to fast fire growth, reaching flashover in 2 min, whilst
a fire having a doubling time of 1 min can be equated to ultra-fast fire growth,
reaching flashover in 1 min. Most of the doubling times obtained in Table 18
were 2 min or under confirming the rapid nature of the fire growth in these extreme
cases.
Table 21
Circumstances surrounding the fires with the highest fire growth parameter (a) values investigated in
occupancy groups found in other buildings
Table 21 (continued)
The estimated fire size doubling time, tdb ; for each incident is also given (where available).
over combustible wall and ceiling linings mean that the flames and hot gas products
released are concentrated on the surface and that the surface ahead of the flame is
pre-heated enhancing the flame spread rate.
circumstances this can produce a runaway feedback reaction and result in very rapid
flame spread along a surface [32].
Table 21 also shows the estimated fire size doubling time, tdb ; for incidents where
the necessary data was available. Most of the doubling times observed in these
extreme cases were under 2 min. The fire doubling times were particularly small in
the cases of a school fire where paint de-lamination caused extremely rapid flame
spread along a corridor and in a bookmakers where a row of TV monitors caught
fire directly under a combustible timber lined ceiling.
6.9. Limitations
One important limitation stems from the limited sample of data available for
analysis, particularly in the case of other buildings, where the data sampled must be
divided over a number of different occupancies. The situation is particularly acute in
the calculation of fire growth parameter distribution values for certain occupancies,
where only a very small number of incidents were available. In such cases the results
obtained must be treated with caution and may be best regarded as being indicative
of the possible magnitude of fire growth parameter that may occur, rather than
providing a definitive value. Hopefully, this situation will become better resolved in
time, as more data from fire investigations becomes available.
The analysis presented here looks only at consequences of fire, examining the
frequency distributions of fire loss and fire growth rates given that a fire has occurred
and been investigated by the fire brigade. However, an assessment of risk must also
take into account the frequency of fire occurrence as well as the consequences of a
fire. An earlier paper by Holborn et al. [33] has considered the frequency of
occurrence of fires in different types of workplace occupancies (in this analysis the
level of fire damage was treated simply in terms of the percentage of fires spreading
beyond the first room). These results may be combined with the consequences
quantified here (in terms of fire damage size) to provide an overall assessment of the
relative risk to a particular occupancy type.
Based upon the fire investigator attendance criteria it might also be argued that the
sample of fires investigated represents the more ‘‘significant’’ (i.e. large damage, high
growth rate) fires that occur and which are attended by London Fire Brigade. The
distributions found here might therefore be regarded as being somewhat pessimistic
in the sense that they do not include the effect of a large number of smaller fires,
which the fire investigators (or even the fire brigade) did not attend.
Morgan [21] has observed that any design based on an average fire growth curve
implies a failure rate of one in two. He therefore suggests that to ensure a reasonable
estimate of the likely fire growth, the fire growth parameter selected to represent a
given design fire should be a specified number of standard deviations above the mean
value of the (fire growth parameter) distribution appropriate to a particular class of
occupancy. As an alternative the x95 value has also been supplied in this analysis
to provide a measure of the location of the tail for both fire damage area and
fire growth parameter distributions. However, it would appear that the x95 location
for the log-normal distributions in many cases underestimate the actual location of
ARTICLE IN PRESS
522 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
the 95th percentile in the original data set, i.e. the tail of actual distribution extends
to larger values than would be predicted by the fitted log-normal distribution.
The current study was limited to London buildings and so the results may not be
directly applicable to some structures in other countries (e.g. Scandinavia, North
America) which have dissimilar construction practices. Indeed, it would be
interesting to compare the results found here, with similar data obtained in other
countries, to examine any similarities and differences in fire behaviour that might be
apparent.
7. Conclusion
Data obtained from fire investigations made by London Fire Brigade between
1996 and 2000 has been used to characterise the distributions of fire damage size, fire
growth rates and the discovery and call times exhibited by fires, in a form suitable for
use with probabilistic risk assessment, in both residential dwellings and other types
of building. The analysis of the results has also highlighted several problem areas.
In residential dwellings there were a number of cases where serious damage was
caused to a property as a result of fires in roof spaces, especially with fires started by
electrical supply and lighting and through the careless use of paint-strippers and
blow lamps.
Extensive damage occurred in a significant number of the warehouse (storage) fires
investigated. Such structures often present a large undivided space for a fire to
spread through. The involvement of sandwich panel cladding and fuel sources with a
fast heat release rate (e.g. plastics) were also found to be important contributing
factors in several cases.
A number of reasons for the occurrence of high growth rates in the fires
investigated were identified including the involvement of flammable vapours and
liquids, thermoplastics, insulation materials in concealed spaces, paint de-lamination
and other extensive distributed fuel sources.
There were also a number of incidents where there was a significant delay between
discovery of a fire and the call to brigade. Reasons found for such delays occurring in
dwellings were that the fire was thought to be extinguished, only for it to be later
discovered that it was still burning or that a neighbour discovered the fire but
delayed calling the brigade, either because they were not sure that it was the correct
thing to do or because they did not know how to make such a call. In other
buildings, there were several cases where previous false alarms meant that a real fire
was not treated seriously and was either initially ignored or had to be investigated
and confirmed before a call to the brigade could be made.
Acknowledgements
The authors would like to express their thanks to London Fire Brigade for their
support in carrying out this research and to all the LFB fire investigators who
ARTICLE IN PRESS
P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524 523
collected data from the scene of fires into the Real Fire Library and made this
analysis possible.
References
[1] Wilmot RTD. Information bulletin of the world fire statistics. Geneva: The Geneva Association;
October 2002.
[2] Ramachandran G. The economics of fire protection. London: E & FN Spon; 1998.
[3] Steiner NR. Lessons from the investigation and analysis of real fires. Ph.D. Thesis, South Bank
University; 1998.
[4] LFCDA. Fire investigation procedures and protocols, operational note 272, London Fire and Civil
Defence Authority, 1996.
[5] Ramachandran G. Extreme value theory and fire losses: further results. Fire Research Note 910, Fire
Research Station, Borehamwood, 1972.
[6] Ramachandran G. Extreme value theory and large fire losses. ASTIN Bulletin 1974;7(3):293–310.
[7] Ramachandran G. Extreme order statistics in large samples from exponential type distributions and
their application to fire loss. In: Patil GP, Kotz S, Ord JK, editors. Statistical distributions in scientific
work, Vol. 2. Dordrecht: D. Reidel; 1975. p. 355–67.
[8] Shpilberg DC. The probability distribution of fire loss amount. Journal of Risk and Insurance
1977;44(1):103–15.
[9] Fontana M, Favre JP, Fetz C. A survey of 40,000 building fires in Switzerland. Fire Safety Journal
1999;32:137–58.
[10] Ramachandran G. Properties of extreme order statistics and their application to fire protection and
insurance problems. Fire Safety Journal 1982;5:59–76.
[11] Ramachandran G. Extreme value theory. In: DiNenno PJ, Beyler CL, Custer RLP, Walton WD,
Watts JM, Drysdale DD, Hall JR, editors. SFPE handbook of fire protection engineering. Quincy,
MA: National Fire Protection Association; 1996.
[12] Vose D. Risk analysis: a quantitative guide. Chichester: Wiley; 2001.
[13] Phillips WGB. Computer simulation for fire protection engineering. In: DiNenno PJ, Beyler CL,
Custer RLP, Walton WD, Watts JM, Drysdale DD, Hall JR, editors. SFPE handbook of fire
protection engineering. Quincy, MA: National Fire Protection Association; 1996.
[14] Rutstein R. The estimation of the fire hazard in different occupancies. Fire Surveyor 1979;8(2):
21–5.
[15] Ramachandran G. Probabilistic approach to fire risk evaluation. Fire Technology 1988;24(3):
204–26.
[16] Ramachandran G. Probability based building design for fire safety, part 2. Fire Technology
1995;31(4):355–68.
[17] Custer RLP. Fire and arson investigation. In: Cote AE, Linville JL, editors. Fire protection
handbook. 18th ed.. Quincy, MA: National Fire Protection Associatio; 1997.
[18] Nelson HE. An engineering analysis of the early stages of fire development—the fire at the Du Pont
Plaza hotel and casino December 31 1986. US National Bureau of Standards, NBSIR 87-3560.
Washington DC, 1987.
[19] NFPA 92B. Guide for smoke management systems in malls, atria and large areas. Quincy, MA:
National Fire Protection Association; 1991.
[20] BSI DD 240. Fire safety engineering in buildings. London: British Standards Institution; 1997.
[21] Morgan HP. Sprinklers and fire safety design. Fire Safety Engineering 1998;5(1):16–20.
[22] Morgan HP, Ghosh BK, Garrad G, Pamlitschka R, De Smedt JC, Schoonbaett LR. Design
methodologies for smoke and heat exhaust ventilation. BR 368, Building Research Establishment,
CRC Limited, 1999.
[23] Ramachandran G, Bengstson S. Exponential model of fire growth. Proceedings of first international
symposium on fire safety science. Washington, USA: Hemisphere Publishing Corporation; 1986.
ARTICLE IN PRESS
524 P.G. Holborn et al. / Fire Safety Journal 39 (2004) 481–524
[24] Wright MS, Archer KHL. Further Development of Risk Assessment Toolkits for the UK Fire
Service: technical note—financial loss model. UK Home Office Technical Note C8155/D99052,
March 1999.
[25] Sardqvist S, Holmstedt G. Correlations between firefighting operations and fire area: analysis of
statistics. Fire Technology 2000;36(2):109–30.
[26] Charters D, McGrail D, Fajemirokun N, Wang Y, Townsend N, Holborn P. Preliminary analysis of
the number of occupants, fire growth, detection times and pre-movement times for probabilistic risk
assessment. 7th International Symposium of Fire Safety Science. 16th—21st June 2002, Worcester,
Massachusetts, USA.
[27] CIBSE Guide E. Fire Engineering. London: The Chartered Institution of Building Services
Engineers; 1997.
[28] Buchanan AH. Structural design for fire safety. Chichester: Wiley; 2001.
[29] Fullwood RR, Hall RE. Probabilistic risk assessment in the nuclear power industry. Oxford
(New York/Sydney): Pergamon Press; 1988.
[30] Conover WJ. Nonparametric statistics. New York: Wiley; 1980.
[31] Hall JR. Use of fire incident data and statistics. In: Cote AE, Linville JL, editors. Fire protection
handbook. 18th ed.. Quincy, MA: National Fire Protection Association; 1997.
[32] Murrell J, Rawlins P. Fire Hazard of Multilayer Paint Surfaces, in Fire and Materials—4th Intl.
Conf. and Exhibition. London: Interscience Communications Ltd; 1995. p. 329–38.
[33] Holborn PG, Nolan PF, Golt J, Townsend N. Fires in workplace premises. Fire Safety Journal
2002;37:303–27.