Content uploaded by Margaret Simonson McNamee
Author content
All content in this area was uploaded by Margaret Simonson McNamee on Nov 12, 2019
Content may be subject to copyright.
*Correspondence to: Dr E. Mikkola, VTT Building Technology, Fire Technology, PO Box 1803, FIN-02044, VTT, Finland
Contract/grant sponsor: Commission of the European Communities
Contract/grant sponsor: Centre for Metrology and Accreditation, Finland
Contract/grant sponsor: EU/FoUradet, Sweden
Contract/grant sponsor: SRV, Sweden
Contract/grant sponsor: Nordtest, Finland and Sweden
FIRE AND MATERIALS
Fire.Mater.24, 101}112 (2000)
Smoke Gas Analysis by Fourier Transform Infrared
Spectroscopy }Summary of the SAFIR Project Results
Tuula Hakkarainen, Esko Mikkola*, Jan Laperre, Francis Gensous, Peter Fardell, Yannick Le Tallec,
Claudio Baiocchi, Keith Paul, Margaret Simonson, Caroline Deleuand Edwin Metcalfe
VTT Building Technology, Fire Technology, PO Box 1803, FIN-02044 VTT, Finland
Centexbel, Technologiepark 7, B-9052 Zwijnaarde, Belgium
ELF Atochem, Groupement de Recherches de Lacq, PDSA, R.N. 117, B.P. 34, F-64170 Lacq, France
Fire Reseach Station, Building Research Establishment, Bucknalls Lane, Garston, Watford, WD2 7JR, United Kingdom
Laboratoire National d'Essais, CEMATE, Division comportement au feu, Z.A. de Trappes-Elancourt, 5 Avenue Enrico Fermi, F-78197
Trappes Cedex, France
L.S.F. SUD srl, Laboratorio di Studi e Ricerche sul Fuoco, Via Boni"ca 4, I-64010 Controguerra (TE), Italy
RAPRA Technology Limited, Shawbury, Shrewsbury, Shropshire, SY4 4NR, United Kingdom
Swedish National Testing and Research Institute, Fire Technology, PO Box 857, S-50115 Boras, Sweden
Laboratory for Heat Transfer and Fuel Technology, University Ghent, Ottergemsesteenweg 711, B-9000 Gent, Belgium
University of Greenwich, School of Chemical and Life Sciences, Woolwich Campus, Wellington Street, Woolwich, London SE18 6PF,
United Kingdom
The determination of toxic components from 5re gases is di7cult because the environment is hot, reactions are often
temperature dependent, and a lot of soot may be produced. Due to the di4erent properties of the gas components,
adi4erent time-consuming procedure for each species has traditionally been used. The use of FTIR (Fourier
transform infrared) spectrometers as a continuous monitoring technique overcomes many of the problems in smoke
gas analyses. FTIR o4ers an opportunity to set up a calibration and prediction method for each gas showing
a characteristic spectral band in the infrared region of the spectrum.
The objective of the SAFIR project was to further develop the FTIR gas analysis of smoke gases to be an applicable
and reliable method for the determination of toxic components in combustion gases related to 5re test conditions. The
optimum probe design, 5lter parameters and the most suitable sampling lines in terms of 6ow rate, diameter,
construction material and operating temperature have been speci5ed. In the large scale, special concern was given to
the probe design and the e4ects of the probe location as well as practical considerations of the sampling line length.
Quantitative calibration and prediction methods have been constructed for di4erent components present in smoke
gases. Recommendations on how to deal with interferents, non-linearities and outliers have been provided and
a veri5cation method for the spectrometer for unexpected variations and for the di4erent models have been described.
FTIR measurement procedures in di4erent 5re test scenarios have been studied using the recommendations of this
project for measurement techniques and analysis and an interlaboratory trial of the FTIR technique in smoke gas
analysis was carried out to de5ne the repeatability and reproducibility of the method in connection with a small scale
5re test method, the cone calorimeter. Copyright 2000 John Wiley & Sons Ltd.
INTRODUCTION
This paper summarises the work performed in the
SAFIR (Smoke Gas Analysis by Fourier Transform
Infrared Spectroscopy) project within the European
Standards, Measurement and Testing programme under
Contract no. SMT4-CT96-2136. The project was in-
itiated in February 1997 by the Commission of the
European Communities, DG XII, in the SMT research
programme.
The members of the SAFIR Consortium were: VTT
Building Technology (Finland), Scienti"c and Technical
Centre of the Belgian Textile Industry (Belgium),
Groupement de Recherches de Lacq (France), Building
Research Establishment/Fire Research Station (UK),
Laboratoire National d'Essais (France), L.S.F. SUD
srl}Laboratorio di Studi e Ricerche sul Fuoco (Italy),
RAPRA Technology Limited (UK), Swedish National
Testing and Research Institute (Sweden), Universiteit
Gent (Belgium), and University of Greenwich (UK). The
coordinator of the project was VTT.
The SAFIR project was divided into "ve work pack-
ages: WP1 *Sampling in small scale; WP2 *Sampling
in large scale; WP3 }Data analysis, calibration and soft-
ware; WP4 *Veri"cation in di!erent "re scenarios; and
WP5 }Interlaboratory trial.
The objectives of the SAFIR project were to further
develop the FTIR gas analysis of smoke gases to be an
applicable and reliable method for the determination of
Received 1 June 1999
Copyright 2000 John Wiley & Sons, Ltd. Accepted 1 March 2000
Table 1. Fuels and combustion products of interest in addition to
CO
2
and CO
Material/compound Acrolein H
2
S HBr HCl HCN NO
x
SO
x
Particle board x x x
FR Vexible PU foam x x x
PVC sheet x
Rubber x x
FR polyethylene x x
toxic components in hot combustion gases related to "re
test conditions; to produce a code of practice for the
FTIR method which includes techniques of sampling,
calibration and analysis; to assess the area of application
of FTIR in di!erent "re test scenarios; to prepare a draft
standard (including repeatability and reproducibility
data) in a form suitable to be forwarded to standards
organisations.
EXPERIMENTAL
Sampling
The basic requirements for the sampling system are: (1)
the "re e%uents shall remain unaltered during their pas-
sage, and (2) the transport of the gases is as quick as
possible. The soot particles shall be mainly "ltered out of
the gas sample before it #ows through the sampling line.
The sampling line parts studied were the probe to collect
e%uents from the "re e%uent stream, the "lter to retain
particles, the transfer line to transport the e%uents to the
FTIR analysis, and the gas cell. In addition, the applica-
bility of the sampling system as a whole was veri"ed. The
special features of large scale sampling were also con-
sidered.
Probe. The study of the probes was concentrated on
linear probes made of stainless steel. The tests were
carried out using the cone calorimeter.The material
used was FR ABS, producing large amounts of smoke
and soot. The e!ects of the hole size, the probe orienta-
tion (i.e. holes downstream or towards the #ow), and the
sampling #ow rate on the blocking of the probe holes
were studied.
Filter parameters. The "re model used for the experi-
mental "lter study was the ASTM E 662 test method.
Two materials were used: plasticised PVC and expanded
rubber. The "lter parameters studied were temperature,
membrane diameter and material, porosity, and "lter
holder material. The shape of the "ltering device was
circular in all experiments.
Sampling line. Di!erent types of gas sampling lines and
conditions for "re gas sampling and FTIR analysis were
studied. Evaluations were made mainly for CO, CO,
HCN and HCl since these gases were considered to be
the most relevant in terms of toxicity and di!erent behav-
iour patterns. The transfer line parameters studied were
the temperature, #ow rate, line material and length.
Gas cell. In the gas cell study, the response times of
di!erent FTIR gas cells were determined. In addition, the
e!ect of the gas cell pressure on the measured concentra-
tion was evaluated.
Large scale experiments. The main emphasis of sampling
studies in the large scale was set on door measurements
because the highest soot concentrations and temper-
atures were expected to be found in the door.
The door experiments were performed with a propane
burner or a heptane pool burning in the ISO 9705 room
and doping the #ames with HCl from a gas cylinder. The
gases coming out of the door were sampled using linear
probes with di!erent hole arrangements. Comparison
samples were taken in the standard duct sampling posi-
tion of the ISO 9705 room for CO(NDIR analyser) and
HCl (impinger bottles).
Calibration, analysis and software
Minimum detection limits. The pathlength of the infra-
red beam in the gas cell and the signal-to-noise ratio
of the detector determine the minimum detection
limits (MDLs) of a FTIR spectrometer. A procedure
to evaluate the MDLs for various gases was de"ned,
and several FTIR spectrometer con"gurations
were evaluated. Both MCT and DTGS detectors were
included in the study.
Model building. The applicability of several multivariate
techniques for smoke spectra analysis was studied. The
work concentrated on two main branches: (1) classical
chemometrical techniques (CCT) including CLS, PLS
and INLR, and (2) quantitative target factor analysis
(QTFA) as an alternative approach. Models were built
for CO, CO, HCl, HCN, NO, NO, acrolein and SO.
Procedures for dealing with interferents and overlapping
peaks were de"ned. In addition, the predictions of the
two analytical approaches were compared.
Software. Software packages for analysing smoke gases
were developed based on both CCT and QTFA. The
programming environment was Matlab 5.0 for QTFA
and Visual Basic 4.0 for CCT. Both programmes were
designed for Windows 3.11 and Windows 95.
Veri5cation and round robin
Fuels and scenarios used for veri5cation. The following "re
types, as de"ned by ISO,were represented in the veri"-
cation experiments: 1. Non-#aming: c. non-oxidative py-
rolysis; 2. Well ventilated #aming; 3. Less well ventilated
#aming: a. small vitiated "res in closed compartments,
and b. post-#ashover "res in large or open compart-
ments.
Materials for combustion were chosen to provide key
chemical species as combustion products, some of which
were known to be problematic in "re gas analysis. The
fuels chosen for the study and the key combustion prod-
ucts of interest in addition to COand CO are presented
in Table 1.
Non-dispersive infrared (NDIR) spectroscopy was
used as a comparison method for COand CO. For HCl,
102 T. HAKKARAINEN E¹A¸.
Copyright 2000 John Wiley & Sons, Ltd. Fire Mater.24, 101}112 (2000)
HBr, HCN and SO, the gas sample was adsorbed in
appropriate solutions and analysed by high performance
ion chromatography (HPIC). Also photometry was used
for HCN. The comparison method for NOVwas chemi-
luminescence. For acrolein, the gas sample was derivitised
on dinitrophenylhydrazine cartridges and analysed by
high pressure liquid chromatography (HPLC).
Round robin and statistical analysis. The interlaboratory
trial was carried out to determine repeatability and re-
producibility of the method according to the ISO 5725
principles.The cone calorimeter methodwas used with
the horizontal specimen orientation and the heat #ux
level of 50 kW/m. The materials tested in three repli-
cates were the following:
1. Particle board, thickness 12 mm, density 700 kg/m;
2. Fire-retarded (FR) PVC, thickness 3 mm, density
1180 kg/m;
3. Polyurethane (PUR) foam panel, thickness 35 mm,
density 40 kg/m.
RESULTS
Sampling in small scale
Probe. A multi-hole probe made of stainless steel was
found suitable for collecting a representative sample of
non-homogeneous gaseous products. The probe shall be
long enough to sample over the whole diameter of the
duct. An open-ended probe can be used if the smoke
gases are mixed uniformly or sampling from a single
point is appropriate.
To avoid the blocking of the probe holes, a minimum
hole diameter of 3 mm is recommended for the FTIR
probe in cone calorimeter experiments. In addition, the
probe holes should be oriented downstream to reduce the
blocking. The blocking tendency of the holes was found
to increase with decreasing #ow rate.
The probe should be cleaned at least in connection
with normal service of the test equipment. If exception-
ally sooty tests are performed, cleaning should be carried
out more often. Cleaning is recommended also between
the test series of di!erent products to avoid disturbance
from possible traces of soot.
In addition to the cone calorimeter tests, the applica-
bility of the results to other small scale test methods
were estimated. The exposed area and thickness of the
specimen, the smoke and soot production, the cross area
of the exhaust duct, and the burning characteristics of
specimens in each test method were considered in the
calculations. According to these estimations, the same
size of minimum probe opening can be recommended for
the DIN furnace method as for the cone calorimeter
method.
Filter. In the study of circular "lters, it was observed that
the amount of HCl retained on the "lter decreases and
the amount of HCl detected by FTIR increases when the
temperature of the "lter is increased.
The stainless steel "lter holder exhibited the best per-
formance in the tests. The PTFE "lter holder was prob-
lematic due to its tendency to seal at high temperatures.
The polyamide "lter holder showed an intermediate be-
haviour.
The "lter shall be dense enough to remove soot par-
ticles from the "re gas sample e$ciently. On the other
hand, the lower the porosity of a "lter is, the more it
obstructs the sampling #ow rate. As a compromise, gen-
eral consideration suggests the choice of a "lter with
5lm porosity.
The results obtained in the comparison of PTFE and
glass "bre membranes with respect to the detectability of
acid gases (HCl and HBr) were divergent for di!erent
materials, depending on the type of the deposit formed.
A disadvantage of PTFE membranes is that they ob-
struct the sampling #ow rate more than glass "bre mem-
branes. On the basis of general aspects and evaluations,
glass "bre membranes are recommended. However, if
a considerable amount of HF is expected to be produced
in the test, PTFE membranes or ceramic "lters are rec-
ommended instead of glass "bre "lters due to the capabil-
ity of HF to etch glass.
The amount of blocked HCl and HBr increases
with increasing "lter diameter. On the other hand, ob-
struction of sampling #ow rate is reduced for larger
"lters. A compromise between blocking of gases and #ow
rate obstruction is a "lter diameter of about 47 mm.
To summarise, the following "lter con"guration is
recommended for circular "lters: holder made of stainless
steel, glass "bre membrane (except for HF), diameter
about 47 mm, and porosity 5 lm. The temperature of the
"lter should be as high as possible with minimum of
1503C and practical maximum of about 1903C.
Independently of its type and shape, the "lter shall be
placed between the probe and the transfer line, as close to
the probe as possible. Cylindrical or circular "lters can be
used depending on the "re test and the amount of soot
expected to be produced during the combustion. The
area of the "lter is probably the most important factor in
the e$ciency.
If acid gases are measured or if there are di$culties
(e.g. blockage) during the test, the "lter should be ana-
lysed after the test to determine the amounts of species of
interest adsorbed to the soot on the "lter.
Sampling line. In the "rst part of the sampling line test
programme, doses of "re gases were passed through 4 m
long gas sampling lines of di!erent materials (PTFE,
PFA and stainless steel) at three temperatures and three
#ow rates, and the response times of the lines were
determined. Most gases behaved in a similar manner
except HCl and HBr which gave signi"cantly longer 90%
response times.
The inner diameter of the sampling line could
be 3}4 mm. The temperature of the line shall be as
high as possible with a minimum of 1503C and practical
maximum of 1903C. The #ow rate of the sample gas in
the line shall be as high as possible with a minimum
of 3.5 l/min.
PTFE and PFA o!ered generally similar performance
which was better than that of stainless steel. Since PFA
presented no signi"cant improvements compared with
PTFE, further tests were carried out with PTFE lines.
In the second phase with PTFE lines, the e!ects of the
line length and pre-ageing were studied.
SMOKE GAS ANALYSIS *SAFIR PROJECT 103
Copyright 2000 John Wiley & Sons, Ltd. Fire Mater.24, 101}112 (2000)
Figure 1. InVuence of gas cell pressure on measured CO
2
con-
centration. The calibration pressure is 650 torr.
PTFE lines of 2 and 10 m showed generally similar
results to those obtained with the 4 m line for CO,CO
and HCN. The results for HCl were di!erent from other
gases and showed longer 90% response times for both
increasing and decreasing edge of a gas dose.
On the basis of the test results, the sampling line should
be as short as possible; 4 m is acceptable. Problems may
arise with longer lines because of greater delays and
sample losses in cold points where condensation can arise.
The adsorption of HCl on the inner surface of the
sampling line decreased when the PTFE line was
pre-aged. Lines which have been previously used for "re
testing o!er improved performance for HCl but it may
be di$cult to obtain uniformity between tests with pre-
aged lines. However, optimum results require the line to
be exposed to "re smoke containing HCl or at least to
HCl before use.
Gas cell. The main factor in the response time is the gas
cell. The delay e!ects of the sampling line and the "lter
are usually of minor importance. The variation of the
response time is not directly proportional to the volume
of the cell and/or the sampling line.
The e!ect of the gas cell pressure on the measured
concentration was evaluated experimentally. The results
obtained showed a signi"cant e!ect on the evaluation of
concentration as illustrated in Fig. 1. The connection
between the absorbance Aand cell pressure pis:
A
A
"p
p
in which Aand pare the absorbance and cell pressure
during calibration. Pressure variation in the gas cell leads
to variation of absorbance and consequently to variation
of the measured concentration. For approximate correc-
tion (non-linear behaviour of COnot taken into ac-
count), the concentrations measured were multiplied by
the ratio of the calibration pressure and the actual pres-
sure, also shown in Fig. 1.
Sampling system as a whole. For the sampling system as
a whole, the following recommendations are given:
EThe point for gas collection in the exhaust duct shall be
placed at a distance where the gaseous mixture is
homogeneous and the gas #ow is not disturbed.
EThe distance between the combustion area and
the probe should be as short as possible to avoid
condensation.
EThe main "lter shall be placed between the probe and
the transfer line. Another "lter can be used between the
line and the gas cell to protect the cell from "ne shoot
particles.
EThe #ow meter shall be placed after the pump. It is
recommended to trap the water between the pump and
the #ow meter.
EThe temperatures of the "lter, sampling line and gas
cell should be as similar as possible. If the same temper-
ature cannot be maintained in all parts of the sampling
system, the temperature shall preferably increase to-
wards the end of the device.
The presence of Cl or Br in unused "lters was checked
by washing the "lters and analysing the solutions.
Non-negligible amounts of chlorine was observed in un-
used (cylindrical) "lters. Therefore, the initial quantity of
Cl or Br in the new "lters must be checked. If these
quantities are high, and if HCl and HBr retained by the
"lter during an experiment must be measured, the "lter
must be washed before use. Washing with hot water
followed by drying in an oven is suitable.
In the veri"cation study of the sampling device, the
observed quantities of HCl and HBr retained by the "lter
were relatively low, 500}1000 lg. Due to saturation after
a certain exposure of acid gases, the HCl and HBr reten-
tion in the sampling device causes no major problems at
high concentration levels.
The response time of the sampling line is longer for
HCl and HBr than for other gases. Therefore, maximum
concentrations will be underestimated especially in case
of fast concentration changes of these gases.
Figure 2 presents a diagram of the optimised sampling
device.
Sampling in large scale
The results of the large scale sampling experiments are
summarised in Table 2. The experiments with the pro-
pane burner exhibited a constant heat release rate and
COproduction across the experimental period. Thus,
the evaluation of COboth in the duct and door is
reasonably reliable. The heptane "res, however, showed
a less stable behaviour, resulting in somewhat less exact
COconcentrations.
The concentration of COwas used to verify the fun-
ctionality of the FTIR measurements in the door. An
average #ow through the door was calculated using
Rockett's theory.Using this information measurements
in the door can be corrected to correspond to expected
results in the duct. These have been compared with
actual measurements made in the duct for CO. As seen
from Table 2, all experiments show an acceptable CO
percentage agreement between the results for the door
and the duct.
The HCl comparison was made between the total
amount of HCl measured in the door and duct due to the
unstable HCl concentrations throughout the experi-
ments. As shown in Table 2, agreement between measure-
ments in the duct and in the door is good when HCl
concentrations have been reasonably high. This is in
104 T. HAKKARAINEN E¹A¸.
Copyright 2000 John Wiley & Sons, Ltd. Fire Mater.24, 101}112 (2000)
Figure 2. Diagram for optimised sampling system.
Table 2. Comparison between CO
2
measurements made in the door and the duct
Fuel
HCl added
(ppm)
a
%CO
2
in duct
Corrected %CO
2
in door
Difference
(%)
HCl
total
in duct
(g)
HCl
total
in door
(g)
Difference
(%)
Propane ca. 1000 0.84 0.87 ;3.5 415 356 914
Propane ca. 100 0.81 0.85 ;4.9 Y
b
Y
b
Y
b
Propane ca. 100 0.83 0.84 ;1.2 14 44 ;214
Propane ca. 1000 0.81 0.81 <0 271 283 ;4
Heptane ca. 100 0.39 0.38 92.6 45 40 911
Heptane ca. 1000 0.42 0.40 94.8 281 241 914
Heptane ca. 100 0.42 0.40 94.8 Y
c
Y
c
Y
c
a
These numbers correspond to desired concentrations. The concentrations measured were generally lower.
b
No comparison due to unreliable impinger bottle measurements.
c
No veriUcation measurement for HCl.
agreement with the "ndings in the study of small scale
sampling which indicate that the system becomes
saturated after a certain exposure to HCl. In the case of
low concentration levels this saturation is not achieved
and the HCl measurements both in the door and in the
duct are unreliable.
The amount of HCl detected throughout the tests was
considerably lower than expected on the basis of the
amount of HCl introduced into the #ames. The probable
reason is HCl loss in the feeder line due to condensation
water trapping HCl.
The retention of HCl in cylindrical "lters was negli-
gible for all experiments. For planar "lters, however,
a slightly larger retention was observed. In all cases, the
percentage retention of HCl was higher when the HCl
concentration was low, indicating saturation behaviour
of HCl adsorption.
In conclusion, most of the recommendations from the
small scale investigations are applicable to the large
scale. There are, however, some features in the measuring
system that are speci"c for large scale testing. When
measuring in the inhomogeneous "re gases the probe
holes should increase in size away from the pump. The
speci"c hole sizes depends on the length of the probe and
hole spacings. Furthermore, it may be di$cult to keep
the sampling line short due to the size of the testing
facilities. In cases where the sampling line is very long
and acid gases are being studied it is advisable to wash
the line after the test to check for adsorbed species.
Calibration, analysis and software
Minimum detection limits. The pathlength and the signal-
to-noise ratio determine the minimum detection limit.
The noise level is a function of the source intensity, the
number of scans, the spectral resolution, type of optical
"lters used, stability of the optical components, and the
quality of the detector. Choosing a low noise detector
such as the MCT detector reduces the noise level. The
noise level of the MCT detector is approximately 100
times less than that of the DTGS detector. Increasing the
number of scans can also reduce noise whereas increasing
the resolution leads to an increase in noise.
To estimate the minimum detection limit of a particu-
lar gas component it is assumed that a spectral line can
be detected when it is approximately as high as the noise
level. We also assume that a linear relationship exists
between the absorbance and the concentration for low
concentrations. Under these assumptions a procedure to
estimate the minimum detection limit for a particular gas
component is as follows:
1. From at least 10 di!erent real smoke spectra deter-
mine the peak-to-peak noise level (pp) of the spectra in
a region free of peaks. Calculate the average 1pp2.
2. Obtain the spectrum of the reference gas with the
lowest concentration (c) and determine the highest
absorbance value (abs ) of the spectrum in the re-
gion you use for building the model.
SMOKE GAS ANALYSIS *SAFIR PROJECT 105
Copyright 2000 John Wiley & Sons, Ltd. Fire Mater.24, 101}112 (2000)
Table 3. An example of the e4ect of resolution and number of
scans on the MDL of CO
Resolution (cm
!1
) Number of MDL
scans (ppm)
2116
48
4115
47
Figure 3. Wavenumber regions with relative absorption activity
of each gas component. (Key: black: absorbance(0,1; light
grey: 0,1(absorbance(1; dark grey: absorbance'1.)
3. The minimum detection limit is given by the ratio of
the peak-to-peak noise level to the highest absorbance
value times the concentration of the reference gas.
MD¸" 1pp2
abs
c
The greater the number of scans the greater the signal-to-
noise ratio. This has an e!ect to the sensitivity of the
spectrometer for di!erent gases. Table 3 shows some
examples of how the combination of resolution and num-
ber of scans in#uences on the MDL of CO.
Model building. In the SAFIR project the applicability of
several multivariate techniques for smoke spectra analy-
sis has been studied. A "rst class of techniques, which we
refer to as classical chemometrical techniques, includes
classical least squares (CLS), partial least squares (PLS)
and implicit non-linear latent variable regression (INLR).
An alternative technique, quantitative target factor anal-
ysis (QTFA) which uses a di!erent approach, has also
been found suitable for gas spectra analysis.
Before discussing the di!erent techniques for analysing
smoke spectra a brief survey of FTIR spectra of the
di!erent toxic smoke gas components (CO, CO,SO
,
acrolein, NO, NO, HCl and HBr) is presented. In Fig. 3
we present the wavenumber regions for each gas where
the absorption peaks appear. From this "gure, it is easily
deduced in which regions absorption peaks of di!erent
gases overlap.
Each gas component exhibits a characteristic FTIR
spectrum. In Fig. 4a}h, a typical reference spectrum of
each gas component is presented.
Water vapour has an infrared spectrum which covers
large parts of the FTIR wavenumber window and com-
plicates the analysis of smoke gases. The FTIR spectrum
of water is presented in Fig. 5.
Recommendations for recording reference spectra. Firstly, se-
lect a number of concentrations of the gas based on
considerations concerning the concentration range for
prediction and the behaviour of the gas (linear or non-
linear). 5 to 6 concentrations for non-linear reference
gases and 3 to 4 concentrations for linear gases is a good
optimum. Buying cylinders with mixtures can be done if
the gases do not have overlapping spectra. Secondly,
start with reliable and accurate recordings of reference
gases. Record the reference spectra with the same experi-
mental settings (path length, resolution, temperature of
the transfer line,2) with which you will record real
smoke gas spectra.
Classical chemometrical techniques (CCT).With classical
chemometrical techniques we denote techniques which
are also applied in other "elds of chemistry, such as CLS,
PLS and INLR.
CLS is a well-known and proven technique for analysis
of FTIR spectra. In this study, the technique was applied
to analyse smoke gases. It was found suitable for
gases behaving linearly, e.g. SOand acrolein. Because
SOand acrolein are covered by water it is necessary to
use gas mixture of water and the gas of interest for
building the model. CLS requires that the concentrations
of water and the gas of interest in the reference spectra
are known. PLS does not have this disadvantage. It only
requires the knowledge of the concentration of the gas
of interest. PLS is unfortunately also a linear technique
and is only suited for linearly behaving gases. Recently,
a modi"ed PLS strategy has been published. This
INLR technique is non-linear and, as shown in this
project, excellently suited for calibration of gases such
as HCl, HCN, CO, CO, NO and NO. Again, the
INLR technique has all the advantages of the PLS
technique.
With classical chemometrical techniques it is possible
to detect outliers. When an unexpected (i.e. not included
in the calibration) and unknown gas component is pres-
ent in the wavenumber window which is used for predic-
tion unreliable predictions will result. F-values, de"ned
as the ratio of the unexplained information during calib-
ration and the unexplained information during predic-
tion, warn for these situations.
There is a number of steps which are very important
for obtaining reliable predictions. These steps are:
Preparing for calibration
1. To avoid over"tting, increase the number of reference
gases to 15 or 30. Each synthetic reference spectrum
should have di!erent concentration. This can be done
numerically (dividing spectra by an appropriate num-
ber) starting from the recorded pure reference spectra.
The concentration of this &&synthetic'' spectrum can be
obtained by a simple area versus concentration model
based on the real reference gases.
2. Several approaches can be followed for obtaining mix-
tures of the reference gas and water.
EThe "rst approach is to obtain water spectra of
di!erent concentrations and mix these spectra
106 T. HAKKARAINEN E¹A¸.
Copyright 2000 John Wiley & Sons, Ltd. Fire Mater.24, 101}112 (2000)
Figure 4. (a)A typical reference spectrum of CO
2
.(b)A typical reference spectrum of CO. (c)A typical reference spectrum of SO
2
.
(d)A typical reference spectrum of Acrolein. (e)A typical reference spectrum of NO. (f)A typical reference spectrum of NO
2
.
(g)A typical reference spectrum of HCl. (h)A typical reference spectrum of HCN.
numerically with the reference spectra. These water
spectra can be obtained by burning methane at dif-
ferent #ow rates. If PLS or INLR is used the concen-
tration of water in these spectra need not be known.
Using these real smoke gas spectra has also another
advantage: introducing noise in the models which
is the same as the noise level during testing. During
the burning of methane also COis generated and
SMOKE GAS ANALYSIS *SAFIR PROJECT 107
Copyright 2000 John Wiley & Sons, Ltd. Fire Mater.24, 101}112 (2000)
Figure 5. FTIR spectrum of water vapour.
therefore these spectra cannot be used for a CO
model. In this case pure water spectra are needed.
ECalibration gas mixtures, including water, can be
purchased ready made in cylinders. These calib-
ration gases can be further mixed or diluted using
suitable equipment to extend the calibration range.
However, experience has shown that particularly for
reactive gases the stated concentration given by the
supplier can be unreliable.
EGas mixtures can also be made in a dynamic system
using permeation or di!usion tubes to produce
a range of concentrations. Gas concentrations are
calculated from the weight loss of the tube. Sub-
sequent analyses for veri"cation can be carried out if
needed.
EA heated, pumped closed loop system attached to the
FTIR cell can also be used to generate known gas
concentrations for calibration. Injecting known
amounts of materials will give known concentrations
if the system is of known volume. After the calib-
ration spectra have been obtained the contents of the
cell and loop can be swept into an appropriate trap-
ping medium for subsequent analysis to determine
concentrations.
Building a model
1. Build your own models and do not rely on other
models because models can be instrument dependent.
2. Start by selecting an appropriate wavenumber win-
dow for calibration and baseline correction. When
selecting a wavenumber window for calibration con-
sider the following:
EAbsorbance should be as high as possible but not too
high (absorbance(1).
EIdeally only the component of interest should show
up in the selected wavenumber region. If this is not
possible, select the window where the component is
covered by water only.
EIf you have knowledge about the wavenumber re-
gions in which unknown interferent gases occur,
avoid these wavenumber windows.
3. Build CLS, PLS and INLR models and select the
number of principal components (PCs) (for PLS and
INLR). In principle di!erent techniques are available
for selecting the number of factors (e.g. cross valida-
tion). Locating the "rst drastic change in the slope of
the standard error of calibration is a good indication
of the number of factors needed. In case there is no
drastic change in slope then use INLR (if you used
a linear technique) and/or reduce the concentration
range. If necessary use cross validation to determine
the number of factors.
4. Select models based on standard error of calibration
and reference versus predicted graphs or other diag-
nostic tools. A standard error of calibration of 10% of
the total concentration range is acceptable.
Predictions
1. Predictions should be checked to be reliable by close
examination of the F-values together with close visual
examination of some spectra.
2. If you "nd suspicious results (e.g. large F-values) and
an unknown gas is identi"ed by visual inspection,
build a new model on a di!erent wavenumber region
not containing the interferent and repeat the predic-
tions.
Quantitative target factor analysis (QTFA).The complexity
of the calibration and modelling process is directly re-
lated to the number of interferents in the system being
modelled. Even if the interferents are not of interest all
the standard multivariate algorithms require that all the
spectroscopically active compounds in a selected
wavenumber region must be present at the time of calib-
ration. If not, the calibration model is not representative.
In such a case a good algorithm should warn the analyst
that the model is incompatible with the sample and that
there are outliers in the sample. If not, the results ob-
tained will be inaccurate. However, the warning or #ag-
ging out of outliers does not solve the problem of
quantifying the target gas or gases. When outliers or
interferences are detected, the whole model has to be
built again in order to ensure reliable quanti"cation. This
can be very time consuming, or even impossible if the
source or identity of the outlier is not known. To circum-
vent some of the calibration di$culties outlined above,
the QTFA approach to "re gas analysis was developed.
Essentially, the QTFA approach estimates the spectral
pro"le of all possible overlapping spectra by a method of
self extraction based on factor analysis which is similar to
principal component analysis (PCA). After estimating the
overlapping spectra, the target gas is quanti"ed with
a simple quadratic absorbance vs concentration curve
deduced from the calibration spectra. The QTFA is
therefore an &open'model, in contrast to PLS which is
a&closed'model.
When applying QTFA, knowledge of PCA is necessary
because the "rst step in QTFA is PCA. QTFA does not
build multivariate models but, after resolving the spectra
(with PCA), uses univariate quadratic calibration models
to predict concentration of the identi"ed target gas(es).
Preparing for Q¹FA
1. The database which is used in QTFA for the identi-
"cation of the di!erent components needs to be con-
structed appropriately.
2. The appropriate wavenumber window should be se-
lected for the gas which is to be analysed.
108 T. HAKKARAINEN E¹A¸.
Copyright 2000 John Wiley & Sons, Ltd. Fire Mater.24, 101}112 (2000)
Figure 7. CO concentration measured by FTIR and comparison
method (NDIR) in cone calorimeter tests of rubber at 50 kW/m
2
.
Figure 6. CO
2
concentration measured by FTIR and comparison
method (NDIR) in cone calorimeter tests of particle board at
50 kW/m
2
.
Q¹FA modelling and prediction
1. After PCA, identify the number of factors using IND
(Factor Indicator Function), %S¸(Percentage Signi"-
cance Level) and PC plots (in the time domain). For
IND, the number of factors (or PCs) is deduced from
the point where the graph passes through the min-
imum. For %S¸, the number of factors is deduced
from the point where there is a drastic change in the
%S¸values (generally, %S¸values between 0 and
5 are signi"cant). For the graphical plots of PCs in
time domain, the visual pro"les with (combustion)
structure are associated with noise. Combining the
information from these three graphs the correct num-
ber of PCs (factors) can be deduced.
2. The number of identi"ed PCs indicates the number of
evolved gases in the selected wavenumber region. If
nPCs are identi"ed, there will be at least 2n!1
estimated spectra generated from which the &KEY'
nspectra are selected. The nmost di!erent spectra
from the 2n!1 initial spectra should be selected
either visually or automatically for the KEY.
3. If any of the KEY nspectra matches the prede"ned
targets in the database then the analysis proceeds with
absolute quanti"cation of the gas(es). Otherwise only
a relative concentration pro"les of the unknown
ngases are reported.
Recommended further reading on the subjects above
has been authored, e.g. by Martens and Naes,
McClure,and Malinowski.
Software. Because QTFA technique is not commercially
available and the calibration strategy of classical
chemometrical techniques cannot be applied without
problems with commercial software, two computer pro-
grammes have been developed to increase the #exibility
and availability of software tools for analysing smoke
gases. The software was written in Matlab 5.0 for QTFA
and in Visual Basic 4.0 for the classical chemometrical
techniques. Both programmes are available for Windows
3.11 and Windows 95.
Conclusions on the calibration and prediction techniques
EFor satisfactory analysis using multivariate or univari-
ate methods, it is essential that good reference spectra
are obtained for calibration. A su$cient number
of calibration spectra should be obtained to allow
modelling of non-linear behaviour with reasonable
precision.
EMultivariate chemometrical techniques such as CLS,
PLS, INLR and QTFA are needed for the prediction of
toxic components in smoke gases generated during
a"re test. For several gases, non-linear techniques are
needed. INLR is non-linear and thus suitable for the
majority of the gases.
EThe QTFA approach is a non-traditional multivariate
method. In QTFA, there is no need for external calib-
ration or mixture preparation and therefore no need
for experimental design of mixtures. The method is
capable of analysing samples from varied sources and
types. The method will be commercially available soon.
EExtrapolation of concentrations outside the calibration
range should be avoided.
EIt is strongly recommended to visually inspect the data
output of the multivariate models. Unexpected behav-
iour (e.g. non-zero baseline) should be investigated.
Veri5cation and round robin
Veri5cation results. In general, the shapes of the concen-
tration vs time curves measured by FTIR were reason-
able considering the "re characteristics of the test. FTIR
gave good peak response, even in the case of rapid
concentration changess. Figures 6 and 7 show example
curves measured for particle board and rubber.
To compare the FTIR results with the comparison
measurements, &FTIR deviation'is de"ned as follows:
FTIR deviation (%)"
FTIR result!Comparison method result
Comparison method result ;100%
The deviations were calculated for total yields and/or
maximum concentrations. There was no evidence to sug-
gest that there was any in#uence of the concentration
level of the gases on the deviation of the FTIR results
from the comparison method.
When the concentration levels measured were low
(near to the minimum detection limit), the relative
SMOKE GAS ANALYSIS *SAFIR PROJECT 109
Copyright 2000 John Wiley & Sons, Ltd. Fire Mater.24, 101}112 (2000)
Table 4. FTIR deviations of CO
2
and CO results compared with
NDIR
Particle FR PVC PUR foam panel
Quantity/Material board (%) (%) (%)
CO
2,max
;4.8 ;8.5 ;7.4
CO
2
yield 90.5 90.6 91.2
CO
max
;8.9 914.1 91.1
CO yield ;16.7 97.8 ;9.3
Table 6. Ranges and average values of relative repeatability and
repoducibility standard deviations for FTIR results
s
r
/
ms
R
/
m
Quantity Range (%) Average (%) Range (%) Average (%)
CO
2
max
4.9}16.2 9.2 14.9}23.8 19.7
CO
2
yield 3.8}15.6 8.3 15.6}39.1 24.2
CO
max
9.9}16.5 13.8 19.8}29.8 25.3
CO yield 5.7}17.0 12.2 34.0}46.7 38.8
HCl
max
a
Y13.0 Y28.5
HCl yield
a
Y8.6 Y26.0
HCN
max
a
Y11.7 Y39.3
HCN yield
a
Y12.2 Y28.7
a
Statistical analysis of HCl and HCN results was performed only
for one material. These single repeatability and reproducibility
values are reported in the ‘‘Average’’ column for HCl and HCN.
Table 5. Ranges and average values of relative repeatability and
repoducibility standard deviations for cone calorimeter quantities
s
r
/
ms
R
/
m
Quantity Range (%) Average (%) Range (%) Average (%)
t
ig
7.3}17.7 12.0 19.0}68.3 36.3
RHR
max
3.9}15.5 8.1 14.3}34.4 22.3
THR 5.6}14.7 8.7 10.2}25.7 16.3
TSP 3.6}14.0 8.7 8.5}17.9 13.1
*h
c
eff
2.4}13.0 6.1 6.2}33.2 16.5
deviation was often exceptionally high. In some experi-
ments, the problems were related to the comparison
method rather than to the FTIR measurements. When
results from several laboratories were available for
a compound, the sign of the deviation seemed to be
laboratory dependent. This suggests that the di!erence
between the results of the methods is due to the measure-
ment procedure of each laboratory, related either to
FTIR or the comparison method. Systematic positive or
negative deviations in all laboratories related to the com-
pound measured or the "re model used were not ob-
served showing that FTIR is a valid method for "re gas
analysis over a wide range of "re test and operating
conditions.
Within the spread of the data there seemed not to be
any particular trend associated with a particular software
package. The scattering of the data presented includes
the possible e!ects of data analysis methods, e.g. those of
the IR spectral regions used in the analyses.
The following presents a summary of the conclusions
for each compound measured.
Good agreement was generally obtained for CO,CO
and NO but the deviations appeared to be laboratory
dependent (either positive or negative). The deviation for
NOwas approximately $50%, but the presence of an
interfering compound in both FTIR and the veri"cation
method make this judgement very uncertain. In HCl and
HCN measurements either good agreement or high pos-
itive deviation values were observed. For acrolein, a devi-
ation of !40% was measured but with too few results to
make "nal conclusions.
In the interlaboratory trial part of the project, using
the cone calorimeter, COand CO concentrations were
measured with both FTIR and NDIR. The FTIR devi-
ations shown in Table 4 were obtained from the mean
values resulting from the statistical analysis.
Summary of veri,cation study conclusions
The recommendations of small and large scale samp-
ling studies were veri"ed. FTIR is capable of making time
resolved measurements on many species simultaneously.
Concentration/time pro"les are generally very good and
realistic and follow the same shape as the comparison
method and the expected pro"le from the "re model.
FTIR gives a good peak response.
The gas cell pressure will in#uence results. If the cell
pressure is not measured and allowed for, the results will
have some uncertainty.
The measured concentrations will be in#uenced by
water vapor. NDIR measurements are with dried gas.
Other comparison measurements are mainly taken at
room temperature, i.e. at a low water vapour pressure.
Not allowing for water vapour will tend to make FTIR
measurements lower than the comparison method.
Repeatability and reproducibility. To analyse the precision
results of the round robin in a proper context it is
necessary "rst to analyse the precision of used "re test
method in relation to the materials or products tested.
Cone calorimeter quantities
The relative repeatability and reproducibility standard
deviations (sP/mand s0/m, respectively) of the cone cal-
orimeter quantities are summarised in Table 5.
Most of these repeatability and reproducibility values can
be considered satisfactory. The main exception is the ignition
time of PUR foam panel. It must be noted, however, that
the mean value of the ignition time for this material was
very small (only 2.6 s), leading to rather high sP/mand s0/m
values. The absolute values of sPand s0for the ignition
time of PUR foam panel were very low, less than 2 s.
In conclusion, the repeatability and reproducibility of the
cone calorimeter results were satisfactory. Thus, the e!ect of
the variability of the cone calorimeter procedures on the
scattering of the FTIR results in this interlaboratory trial
can be considered typical for well de"ned "re tests.
F¹IR quantities
The relative repeatability and reproducibility standard
deviations (sP/mand s0/m, respectively) of the FTIR re-
sults are summarised in Table 6.
Calculated over all materials and gases included in the
statistical analysis, the averages of sP/mand s0/mfor
FTIR results are 11.7% and 25.4% for the maximum
concentration values, and 10.3% and 30.4% for the
yields, respectively.
The FTIR results clearly deviating from the typical
magnitude of the values in Table 6 are the relative repro-
ducibilities of CO yield and HCN. In the statistical
110 T. HAKKARAINEN E¹A¸.
Copyright 2000 John Wiley & Sons, Ltd. Fire Mater.24, 101}112 (2000)
Table 7. Comparison of accuracy data of a selection of test methods
Method/parameters Repeatability standard Reproducibility standard
deviation
s
r
/
m
(%) deviation
s
R
/
m
(%)
SBI, 1997
RHR
max
, THR 5}41 7}60
Smoke production 7}78 19}114
Cone calorimeter, ISO TC92"SC1, 1997
RHR
max
, THR 3}55 4}87
Cone calorimeter, ISO TC92 SC1"WG5, 1991}1992
Smoke production 6}60 16}100
This study
Max. gas concentration and gas yield (FTIR)
a
4}17 15}47
(11.0% in average) (27.9% in average)
RHR
max
, THR 4}16 10}34
Ignition 7}18 19}68
b
Smoke production 4}14 9}18
a
Including results for CO
2
, CO, HCl and HCN.
b
The high maximum value of
s
R
/
m
(68%) is due to the short ignition time of PUR foam panel.
analysis of the CO yields, no outliers or even stragglers
were detected. However, if the results of the laboratory
with the greatest deviation from the mean value were
excluded for each material (not the same laboratory in all
cases), the averages s0/mof CO yield would be reduced to
30.7%. The high s0/mvalue of HCN can be explained
by the relatively low mean value, 40 ppm.
The repeatability and reproducibility values of gas
concentrations and gas yields are best comparable to
those of heat release/total heat release measurements. To
give an indication for the interpretation of the data of this
study, results from previous interlaboratory trials are
reviewed. Table 7 summarises the comparisons with the
results of the SBI and cone calorimeter methods.
The following conclusions can be made on the basis of
the results of Table 7:
Erepeatability standard deviation values for FTIR
measurements are similar to those of RHR and
THR measurements;
Ereproducibility standard deviation values are larger for
FTIR measurements than for RHR and THR
measurements; caused by the scatter in ignition and
burning processes in the test method.
Considering the e!ect of the scattering of the "re test
method included in the FTIR variation, the repeatability
and reproducibility of the FTIR analysis of smoke gases
can be regarded as satisfactory.
CONCLUSIONS
FTIR spectrometers o!er a continuous monitoring tech-
nique of many gases simultaneously in smoke gas analy-
sis. Using FTIR, it is possible to set up a calibration and
prediction method for each gas showing a characteristic
spectral band in the infrared region of the spectrum. The
main problem in the measurements is to have a represen-
tative sample of "re gases analysed.
The principle in de"ning an optimum sampling system
for FTIR was to transfer the "re e%uents through the
sampling device as quickly as possible and keep them
unaltered during the passage. To ensure the representa-
tive smoke gas samples for the analysis, the following
recommendations applicable to both small and large
scale should be taken into account:
EA multi-hole probe made of stainless steel with
minimum hole diameter of 3 mm is recommended.
The holes shall be oriented downstream from the "re
gases.
EBoth circular (glass "bre except for HF) and cylindrical
(e.g. ceramic) "lters can be used. The initial amount of
Cl and Br in the "lters shall be checked.
EA gas sampling line made of PTFE with inner diameter
of 3}4mmo!ers the best performance. The line should
be as short as possible; lengths up to 4 m are found
acceptable.
EThe #ow rate shall be as possible, at least 3.5 l/min.
EThetemperatureofthesamplingdeviceshallbeashighas
possible with minimum of 1503Candpracticalmaximum
of 1903C. To avoid cold points where condensation
could appear, the temperatures of the di!erent parts
should be as close to each other as possible. If the same
temperature cannot be maintained in all parts of the
sampling system, the temperature shall preferably in-
crease towards the end of the device.
EWhen acid gases are analysed by FTIR, it is recommen-
ded that the di!erent parts of the sampling device
should be washed, especially when the concentrations
determined are low. The washing solution should be
analysed by an appropriate analytical method to evalu-
ate the total amount of the gas produced by the com-
bustion.
EThe pressure in the gas cell must be checked during the
experiments because the variations of the pressure in-
side the sampling con"guration leads to a signi"cant
variation in the measured concentrations of gaseous
products.
There are some features of the measuring system that
are speci"c for large scale testing. Making measurements
e.g. in the door of a "re room places extra constraints on
the measuring system.
SMOKE GAS ANALYSIS *SAFIR PROJECT 111
Copyright 2000 John Wiley & Sons, Ltd. Fire Mater.24, 101}112 (2000)
EWhen measuring gases that are not well mixed the
multi-hole probe sampling from the "re gases should
have holes that are graded in size according to where
they are on the probe.
EDue to the size of the testing facilities it can sometimes
be di$cult to keep the sampling line short. In cases
where the sampling line is very long and acid gases are
being studied it is advisable to wash the line to check
for adsorbed acids.
The "ndings of this project concerning sampling tech-
niques are most relevant for all other methods used for
de"ning "re gases.
Quantitative calibration and prediction methods have
been constructed for di!erent components present in smoke
gases. The following conclusions and recommendations are
given related to data analysis, calibration and software:
EMultivariate chemometrical techniques such as CLS,
PLS, INLR and QTFA are needed for the prediction of
toxic components in smoke gases generated during
a"re test. For several gases, linear techniques are
inadequate and non-linear techniques are needed.
EFor satisfactory analysis using multivariate or
univariate methods, it is essential that good reference
spectra are obtained for calibration. A su$cient num-
ber of calibration spectra should be obtained to allow
modelling of non-linear behaviour with reasonable
precision.
EExtrapolation of concentrations outside the calibration
range should be avoided.
EIt is strongly recommended to visually inspect the data
output of the multivariate models. Unexpected behav-
iour (e.g. non-zero baseline) should be investigated.
In the veri"cation of FTIR measurement procedures in
di!erent "re test scenarios, the following conclusion have
been made:
EFTIR is capable of making time resolved measure-
ments on many species simultaneously.
EConcentration/time pro"les are realistic and follow
the shape of the pro"les from comparison methods and
"re test characteristics. FTIR gives a good peak
response.
ECell pressure shall be monitored due to its in#uence on
the concentration results.
EConcentrations will be in#uenced by water vapour, i.e.
the FTIR results can be di!erent from the comparison
method results obtained with dried gas sample.
In an interlaboratory trial using the cone calorimeter
method and testing three materials in three replicates the
FTIR method of measuring smoke gases was found re-
peatable and reproducible when analysed statistically
according to the ISO 5725 principles and compared with
corresponding data of well-known "re test methods. This
precision information obtained has been included in
a proposal for draft standard of the FTIR method for
smoke gas analysis.
Acknowledgements
The Commission of the European Communities contributed to the
SAFIR project through the SMT research programme. The remaining
funding was raised nationally by the partners of the Consortium.
A considerable portion of the work was funded by the partner organisa-
tions. Additional national sponsors were the Centre for Metrology and
Accreditation in Finland, EU/FoUrasdet and SRV in Sweden, and
Nordtest both in Finland and in Sweden.
REFERENCES
1. ISO 5660-1
Fire Tests
_
Reaction to Fire
_
Part
1
:
Rate of Heat
Release from Building Products
(
Cone calorimeter method
)
.
ISO: Geneva, 1993.
2. ASTM E662-97.
Standard Test Method for Speci
[
c Optical
Density of Smoke Generated by Solid Materials.
ASTM: Phil-
adelphia, 1997.
3. ISO 9705
Fire Tests
_
Full-scale Room Test for Surface Prod-
ucts
. ISO: Geneva, 1993.
4. ISO TR 9122-1.
Toxicity Testing of Fire Ef
\
uents
_
Part
1
:
General.
ISO: Geneva 1989.
5. ISO 5725-1
Accuracy
(
Trueness and Precision
)
of Measure-
ment Methods and Results
_
Parts
1
}
6
. ISO: Geneva, 1994.
6. Drysdale D.
An Introduction to Fire Dynamics,
2
nd
edn. John
Wiley & Sons; Chichester, 1998; 307.
7. Martens H, Naes T.
Multivariate Calibration,
John Wiley
& Sons: Chichester, 1989.
8. McClure G
Computerized Quantitative Infrared Analysis.
ASTM: Philadelphia, 1987.
9. Malinowski ER.
Factor Analysis in Chemistry,
2nd edn, John
Wiley & Sons: Chichester, 1991.
112 T. HAKKARAINEN E¹A¸.
Copyright 2000 John Wiley & Sons, Ltd. Fire Mater.24, 101}112 (2000)
