I:resenius leitsehrih [fir

Fresenius Z Anal Chem (1989) 335: 728- 737

© Springer-Verlag1989

Possibilities and limitations of the photoelectric aerosol sensor array applied to heavy metal aerosols* Reinhard Niessner, Beate Hemmerich, and Ulrich Panne University of Dortmund, Inorganic and Analytical Chemistry, P. O. Box 500500, D-4600 Dortmund 50, Federal Republic of Germany

M6glichkeiten und Grenzen des photoelektrischen Aerosolsensors fiir die Analyse yon Schwermetallaerosolen Summary. The technique of substance-selective aerosol charging by photoelectron emission was applied to several ultrafine heavy metal aerosols. To enhance the selectivity of the Photoelectric Aerosol Sensor (PAS), an array operating with four different wavelengths (185 nm, 214 nm, 229 nm, 254 nm) was used. The photoelectric activity at each wavelength and the linearity of the PAS signal was investigated for monodisperse and quasi-polydisperse systems. The aerosol mass concentration was determined by correlation with atomic absorption spectroscopy measurements of filter samples. Pattern recognition was applied to differentiate between four different aerosol systems and to identify common interferences.

1

Introduction

Industrial processes such as welding, smelting or processing of molten metals produce ultra fine metal oxide particles at concentrations, which often exceed in enclosed areas the recommended M A K (maximal allowable concentrations) and T R K (technical guiding concentrations) values [ 1 - 3]. Inhalation of these metal oxides, especially zinc oxide, cadmium oxide, cobalt oxide, and copper oxide are known to cause complaints in the respiratory duct, the gastro-enteric duct, the heart-circulatory system, and the nervous system [1, 4, 5]. As the amounts of welding pollutants emitted into the atmosphere are quite small, the most problems arise near the emission source, where particle diameters are much smaller than 2 btm [4, 6]. Hence it is of considerable interest to use an analytical method with the ability of on line and in situ detection of ultrafine heavy metal aerosols. For welding fume analysis standard aerosol sampling techniques are used like filter sampling with a personal sampler. Gravimetric determination and subsequent atomic absorption spectroscopy (AAS), wavelength dispersive Xray fluorescence spectrometry (WD-XRF) or inductively coupled plasma atomic emission spectroscopy (ICP-AES) [7, 10] are commonly applied to analyse the filter samples. * Dedicated to Prof. Dr. G. T61g on the occasion of his 60th birthday Offprint requests to: R. Niessner

To achieve a representative sample during a working period a sampling time of 5 to 8 h must be considered [4, 8 - 10]. A major drawback is that all these techniques can only be used off line. Alternative approaches are able to detect the total suspended matter (TSP) of particles in the air, but without any speciation [3, 8, 10]. The aim of this study was to explore the possibilities of substance-selective aerosol charging by photoelectron emission for analysis of ultrafine heavy metal particle systems and on line control of occupational exposure.

2

Experimental section

2.1 Reagents

The investigated metals were selected according to their carcinogenicity and mutagenicity, respectively their environmental importance. All metal salts used were of analytical grade and purchased from Merck (Darmstadt, FRG), or Fluka AG (Neu-Ulm, FRG). The metal wires used in the sparking generator were of high purity (>99.999%) delivered from Zinsser Analytik (Frankfurt a.M., FRG). Nitric acid (Suprapur), hydrochloric acid (Suprapur) and the standard metal solutions for AAS analytics were purchased from Merck (Darmstadt, FRG). 2.2 Generation of mono- and polydisperse test aerosols

Aerosols of metal oxides were generated by nebulization of aqueous solutions of metal formates or acetates (c = 1 0 - 2 tool. 1-1, Q = 240 1 • h-a), leading the spray aerosol through a diffusion dryer with subsequent thermal vaporization and chemical degradation to the oxides (Fig. 1 a) [11, 12]. Further metal oxide aerosols were produced with a home-made sparking generator, whose electrodes consisted of high purity wires (~Z~ = 1 mm, c > 99.999%, Q = 150 1 • h -1) of the metal under study (Fig. I b) [13]. Concentrations of the particle systems under study were in the range between 5 - 1 0 2 and 1 • 104 particles per cm-3, the investigated particle diameters were between 15 nm and 100 nm. To identify the generated species of metal oxides at the particle surface all aerosols were investigated by ESCA (XPS) (Model "ESCA Mark 3", Vacuum Generators, East Grinstead, UK). Samples of monodisperse aerosols for electron microscopy were collected on Nuclepore polycarbonate membrane filters ( ~ = 37 ram, pore size = 0.2 gin). These samples were examined and photographed in a Scanning Electron Microscope (SEM) (Model Stereoscan

729

IAeroso[ ~S3 Excess Qir

A

M

o = 2 4 o , i ~

t:~

°

~ .J;~i

~Metol wire

Q=I/+40[/h ~

I

~

Polydisperse ~eroso[

H

i ~...,

Teflon~

Hixing chamber

i

A Q=2L,0[/

|

M

I

DiLutionair 0=1320ffh

3,3M

P

£ ~ [__~k~__J

~ I O= 2-4I/rain

H

A

•M

a

Fig. la, b. Experimental setup for generation of monodisperse aerosols. A Collison atomizer; D diffusion dryer; H heated quartz tube; DMA differential mobility analyzer; S sheath air; E excess air; P polydisperse aerosol; M monodisperse aerosol, b Sparking generator for generation of polydisperse metal oxide aerosols

PAS-Array

Aerosol=

~

~

Condensation Nudeus

Counfer

A,0

Computer IBM/XT

360; Cambridge Instruments, UK). All particles were of spherical shape and no agglomeration was observed. Fractions of monodisperse aerosols were filtered out from the different polydisperse aerosols with electrostatic classifiers (Thermo Systems Instruments (TSI), Inc., St. Paul, MN, Model 3071). To generate artificial, homogeneous, quasi-polydisperse aerosols monodisperse aerosols of different particle sizes were mixed in a ring-gap mixing nozzle. Afterwards this mixed aerosol was diluted with dry and filtered air. As agglomeration of the monodisperse aerosols could jeopardize the measurements, the particle distribution after mixing was determined with a screen-type diffusion battery (TSI, Model 3040) and a condensation nucleus counter (CNC) (TSI, Model 3020). The data were transformed to the actual distribution with an algorithm proposed by Twomey [14-17]. In experiments with binary and ternary mixtures of monodisperse aerosols no agglomeration of particles was observed, which was also confirmed by the SEM studies mentioned above. The calculated distributions were approximately the sum of the monodisperse aerosols. 2.3 Photoelectric aerosol sensor array "PAS-array)

The Photoelectric Aerosol Sensor (PAS) has been described earlier for detection of PAH coated particles [18-21], but

Fig. 2

The photoelectrical aerosol sensor (PAS) array

is also capable to detect other photoemitting materials as the investigated metal oxide particle systems. In this study the selectivity of the PAS was enhanced by using a sensor array operating at four different wavelengths (185 rim, 214 nm, 229 nm, 254 nm). In order to measure the amount of photoelectric charge as a substance relevant signal precharged particles have to be neutralized. Therefore the aerosol was led through a neutralizer containing a Kr-85 source (3.7 • 108 Bq) (TSI). Remaining charged particles were removed in a home-made electrostatic precipitator. Thus the total number concentration of particles were determined by a CNC. The neutralized aerosol stream was split into four equal streams and each was illuminated separately in an irradiation cell. Due to their work function the metal oxide particles were charged by irradiation with UV-light. The charged particles were continuously recorded in four aerosol electrometers (TSI, Model 3068), while the current of each electrometer was transferred on line via a 12 bit A/D converter (Data Translation, Model DT 2814, Marlboro, MA) to an IBM PS/2 Model 30, which was also used for further signal processing. Figure 2 shows the experimental setup of the PAS array. The investigated submicron aerosols and their capability to photoemission are listed in Table 1. The primary decomposition products are dependent from the residence time in the heated quartz tube (Fig. I a) and the decomposition temperature. In particular the studied Ni-compounds

730

Table 1. Photoelectric activity of investigated compounds Substance

Cd-Acetate Cd-Nitrate Co-Acetate Cr-Acetate Cu-Acetate Cu(II)Oxide Mn-Acetate Ni-Acetate Ni-Formate Ni(II)Oxide Pb-Acetate Pb-Nitrate Zn-Acetate Ti(IV)Oxide Fe(III)Oxide Diesel exhaust Cigarette smoke

Decomposition temp. [° C] 760 780 450 870 530 340 560 430 750 770 850 --

Primary decomposition product

Photoelectrical activity [nm] 185

214

229

254

CdO CdO CoO Cr203 Cu20 CuO MnO2 NiO NP' b NiO PbO PbO ZnO TiO2 Fe203

+ + + + + . + + + + + + . . . + +

+ +

+ +

+

+ + --

+ -

+ -

-

.

.

.

+ + + + . . .

. . . + -

. . .

Particle surface is NiO, while bulk material is probably Ni [11] b Ni-formate was used throughout in this investigation for generation of NiO

a

showed significant differences in their photoelectric activity at the four wavelengths. This could be due to an incomplete decomposition and f o r m a t i o n o f non-stoichiometric oxides. The E S C A investigations are not sufficient for a complete chemical characterization o f the particles, which includes a surface and bulk analysis o f the particles. F o r precise information about the decomposition products further studies have to be made.

a p p r o a c h gave the best interpretation of the relationship between the objects. F o r the K N N m e t h o d the number o f nearest neighbours k was determined by optimization, i.e. the prediction ability was tested for different values o f k; in this case k = 3 yielded the optimal prediction. Validation o f the training set was done by the leave-one-out procedure, a correction for different class size in the training set was not applied.

2.4 Statistical methodology

2.5 Filter sampling for atomic absorption spectroscopy

All d a t a for regression were mean centered and scaled to unit variance to remove differences in magnitude. Simple and multiple linear regression ( M L R ) was performed on an I B M microcomputer. F o r all regressions the standard statistics as r a values and standard error on regression were computed, while residual plots and graphs o f predicted vs. actual values were calculated for visual inspection o f regression results. As it is beyond the scope of this p a p e r to describe the applied techniques o f multivariate d a t a analysis and calibration, the reader is referred to the s t a n d a r d literature on chemometrics [22, 23]. F o r e x p l o r a t o r y d a t a analysis principal component analysis (PCA) was applied to the column mean centered and variance scaled data. Principal components were calculated with the N I P A L S algorithm described by H. Wold [24, 25]. Partial least squares regression (PLS) was done using the P L S I algorithm with cross validation [ 2 6 - 2 9 ] . F o r PLS the residual statistics, factor scores and loadings were calculated for a better interpretation o f the data and to check for possible outliers [29]. All signal vectors for cluster analysis [42] and the kNearest N e i g h b o u r ( K N N ) [30, 31] m e t h o d were scaled to unit length, as scaling to unit concentration is not possible for u n k n o w n samples. F o r b o t h methods the d a t a matrix was column mean centered and the Euclidian distance was applied as a measure of similarity. A l t h o u g h for cluster analysis a number of hierarchical methods were tested only the d e n d o g r a m for W a r d ' s m e t h o d is displayed as this

The total a m o u n t of metal oxides species in the aerosol system was independently determined with atomic absorption spectroscopy (AAS). The particles were collected on a Nuclepore p o l y c a r b o n a t e m e m b r a n e filter. The filter was placed parallel to the PAS a r r a y as depicted in Fig. 2. Sampling was done with a flow rate of 1.51. m i n - 1, while the PAS signal was recorded simultaneously. Sampling time was between 5 rain (CdO) and 20 rain ((Ni/NiO) and was dependent on particle concentration, particle size and the elemental specific sensitivity o f the A A S instrument. After sampling the filters were treated with 200 gl conc. nitric acid resp. hydrochloric acid (CdO, CoO, C u 2 0 ) resp. HNO3/HC1 (aqua regia) 1:3 (NiO) for 24 h (CdO, CoO, CuzO) and 48 h (NiO), ultrasonic agitated for 30 min and a p p r o p r i a t e diluted to a total volume o f 1 ml. The extracts were centrifuged and the amounts o f cadmium, cobalt, copper and nickel were determined by graphite furnace atomic a b s o r p t i o n spectroscopy ( G F AAS). A Perkin-Elmer M o d e l 2380 atomic a b s o r p t i o n spectrometer equipped with a deuterium arc b a c k g r o u n d corrector, a Perkin Elmer M o d e l H G A 500 graphite furnace atomizer and an AS 40 autosampler was used.

3 Substance-selective aerosol charging by photoelectron emission W i t h the assumption o f an a p p r o x i m a t e l y homogeneous surface the charging rate o f metal and metal oxide particles

731 under illumination is a function of the following parameters

[32, 33]: dN F + -'f { d rt"t d2 "

Y(h.v),

(1)

~b,v, N }

N÷

= Number concentration of positive charged particles [cm- 3j N = Total number concentration of particles [cm - 31 t = Time [s] 7~ • d z = Irradiated particle cross section [nm 2] F = Fraction of photoemitting material (here assumed as F = 1) Y(h. v)= Photoelectric quantum yield at photon energy h • v (material and wavelength dependent) q~,v = Radiant flux [W. cm-2] Under the applied experimental conditions, i.e. for pure metal and metal oxide particles, the following parameters were kept constant:

F, Y(h. v), N, ~,v, t = const. The number of charged particles and hence the measured signal is directly proportional to the total photoelectrically active surface, which is dependent on the individual particle surface and the number concentration of these particles. N + ~ Stota 1 (7~' d 2, Ni)

(2)

S;.k = Signal of the PAS sensor at wavelength 2k bz~ = Slope at wavelength Zk, characteristic for each compound Ni = Number concentration of particles with diameter d~ di = Particle diameter = Residual To confirm this model monodisperse aerosols of different particle diameters and number concentrations were generated for all investigated species. SEM investigations of these aerosols exhibited particles of approximately spherical shape, hence the total particle surface per unit volume was calculated throughout with the assumption of a spherical shape. The total photoelectrically active surface (always referred to unit volume), calculated as product of the squared particle diameter, ~ and the number concentration, was fitted by least squares to the observed PAS signal. Table 2 demonstrates the excellent correspondence between the postulated model and the experimental results. In Fig. 3 the linear relation between the PAS signal and the total photoelectrically active surface is displayed for monodisperse NiO aerosols at all four wavelengths -~-kFor each system under study artificial, quasi-polydisperse aerosols were used to check, whether Eq. (2) holds also for real polydisperse aerosols. These quasi-polydisperse aerosols were generated by mixing of two or three monodisperse aerosols of different particle diameters and number concentrations. For polydisperse aerosols Eq. (3) is modified to 0o

N+

= Number concentration of positive charged particles [cm- 3] S, ot,~ = Total photoelectrically active surface d~ = Particle diameter N~ = Number concentration of particles with diameter d~

(4)

Sx k = ~b~ k ~ ' f ( d ) d d + 0

or for a mixture of n monodisperse aerosols

(5)

Sx k = b~.~ L re" Nijd~ + j-1

4

Results and discussion

According to theory [Eq. (2)] the PAS signal for monodisperse aerosols should obey a simple linear model of the form

S~k = bzk. N~Tc.d~ + 8

(3)

S ~ = PAS signal at wavelength 2k bx~ = Slope at wavelength 2k, characteristic for each compound f ( d ) = Distribution of particle diameter Nij = Number concentration of particles with diameter i from the monodisperse aerosol stream j" dij = Particle diameter of aerosol s t r e a m j 8 = Residual

Table 2. Results for the linear regression of the total photoelectrically active surface of monodisperse aerosols on the PAS signal Compound

Diameter range ~ [nm]

Data points

2i [nm]

ba~b [nm2 cm 3 mV]

r2¢

CdO CoO

1 6 - 60 10-100

12 22

NiO

1 0 - 70

13

185 185 214 229 185 214 229 254

22.40 ± 3.15 ± 2.52 ± 1.19 ± 10.11 ± 11.17 ± 9.37 ± 6.44 ±

0.995 0.993 0.998 0.991 0.994 0.994 0.996 0.997

Particle diameter was varied within this range b Slope, significant at a level ~(2) = 0.05 ° Coefficient of determination, significant at a level c~(2)= 0.05

0.46 0.06 0.03 0.02 0.23 0.25 0.16 0.10

732 F o r all m i x t u r e s the P A S signal o f e a c h single aerosol stream j was m e a s u r e d , a n d if Eq. (5) holds within the e x p e r i m e n t a l error, the actual P A S signal o f the m i x t u r e should be predictable f r o m the s u m o f these signals. F o r

simplicity this a s s u m p t i o n was checked with a linear m o d e l o f the f o r m StOtal 2k

=

axk

"

~ S{ k + ~

(6)

j=l S t OAk tal

180.00 D

160.00 140.00

:z tz m,.

120.00

illl~ ,IL A

•

= •

100.00

• ~

80.00

•

aa~ Sz~

xx

e

x x A

60.00

x

x

40.00 x

20.00

o%;

I

2.00

4-.00

I

6.00

I

8,00

I

10.00

I

12.00

1

14.00

16.00

Total photoelectrically tic five surfoce per unit votume [nm~/cm a -106]

Fig. 3. Linear relationship between total photoelectrically active surface of monodisperse NiO aerosols and the PAS signal (2k: [] 185 rim, • 214 n m , • 229 nm, x 254 rim; particle diameter between 1 0 - 70 nm)

P A S signal o f t h e m i x t u r e o f j m o n o d i s p e r s e aerosols at w a v e l e n g t h 2k -- Slope at w a v e l e n g t h 2k = P A S signal o f m o n o d i s p e r s e a e r o s o l j at w a v e l e n g t h 2k = Residual

F o r v a r i o u s b i n a r y and t e r n a r y m i x t u r e s o f each system Table 3 a a n d 3 b s h o w the results o f the linear regression. T h e coefficient o f d e t e r m i n a t i o n r 2 indicates a g o o d agreem e n t w i t h theory. N a t u r a l l y the linear additivity o f the P A S signals holds also for m i x t u r e s o f different species as illustrated by Fig. 4 (for clearness the P A S signal is depicted in Figs. 4 a n d 5 only at 2k = 185 nm). A striking p h e n o m e n o n was t h a t the o b s e r v e d P A S signals o f the quasipolydisperse aerosols were nearly in every case b e t w e e n 2 10% h i g h e r t h a n actually p r e d i c t e d (Fig. 5). T h e r e f o r e the slope aa k was significantly different f r o m the expected slope 1.00. A l t h o u g h up to n o w we h a v e n o e x p l a n a t i o n o f this

Table 3 a. Linear additivity of PAS signals for binary mixtures of monodisperse aerosols Compound

Diameter range a [nm]

Data points

2i wavelength [nm]

a~ b

r2~

Slope d

CdO

18-62 62-18 2 0 - 70 90-40

12

185

1.29 ± 0.04

0.998

+

10

20 - 8 5 96-25

13

185 214 229 185 214 229 254

1.01 1.05 1.01 1.03 1.09 1.00 1.04

0.999 0.999 0.998 0.999 0.999 0.999 0.999

-+ -+ + +

CoO

NiO

a b c a

± 0.02 __+0.02 ± 0.03 ± 0.01 ± 0.01 ± 0.01 ± 0.02

Particle diameters were varied within this intervalls Slope, significant at a level cq2) = 0.05 Coefficient of determination, significant at a level c~{2) = 0.05 + Indicates that the slope is significant different from expected slope of t.00 (cq2) = 0.05)

Table 3 b. Linear additivity of PAS signals for ternary mixtures of monodisperse aerosols Compounds

Diameter range a [nm]

Data points

2i wavelength [nm]

aa~ b

r2 c

Slope d

CdO

20-- 50 50 96-- 50 20-80 50 9 6 - 50 2 0 - 50 50 96 - 50

8

185

1.10 _+ 0.02

0.999

+

8

185 214 229 185 214 229 254

0.99 1.08 1.07 1.04 1.10 1.02 1.05

0.999 0.999 0.998 0.999 0.999 0.999

+ + + + -

0.999

+

CoO

NiO

6

± ± + + + + +

0.02 0.01 0.01 0.02 0.02 0.02 0.05

" Particle diameters were varied within this intervalls b Slope, significant at a level c~{2) = 0.05 Coefficient of determination, significant at a level e(2) = 0.05 d ÷ Indicates that the slope is significant different from expected slope of 1.00 (c~(2) = 0.05)

733 310.00

"g

2800.00

link

270.00 2400.00

2.50.00 c 190.00

E

2000.00

i

f

m

rl~

~z o_

c~

15o.oo

1600.00 1200.00

110.00 o._

70.00

30.

I

800.00

~A I

05(~.'00 ' 70.00

~

1

~

110.00

I

I

I

I

~190.00 ~

150.00

2.50.00

, 270.00

fl.

Predicted PAS signa[ ImV]

m ~

400.00

510.00

W

rm l ~

m

, , ,

o%0o

1.00

Fig. 4. Measured and predicted PAS signal for various CoO/NiO mixtures at "¢k = 185 nm

Mass

2.00

concentration

, 4.00

3.00

5.00

I

6.00

[]ag/m 31

Fig. 6. Correlation of the mass concentration and the product of PAS signal and particle diameter for monodisperse CoO aerosols (2k: • 214 nm, [] 229 nm; particle diameter between 2 0 - 9 0 nm) 2700.00

O C3 £3[:3

2500.00

co

19oo.oo

C = a~k S ~ ~f ( d ) dd + e

A

0_ 1500.00 c~

1100.00 700.~I00

'

I

1100.00

,

I

1500.00

(7)

0

A£3

h

I

1900.00

2

I

:,'300.00

,

I

2700.00

C axk S~ f(d)

= = = = =

Mass concentration Slope at wavelength )~k PAS signal at wavelength Particle size distribution Residual

•k

Predicted PAS signat [rnV]

Fig. 5. Measured and predicted PAS signal for artificial, ternary mixtures ofmonodisperse aerosol systems (~ CoO, • NiO and • CdO) at 2k = 185 nm

synergistic effect, on the whole this deviation is no severe violation of the proposed model. 4.1 Correlation between aerosol mass concentration and P A S signal To get an impression of the correlation between the PAS signal and the mass concentration of the heavy metal aerosols, filter samples of monodisperse aerosols were collected for AAS measurements, while simultaneously the PAS signal was recorded. The PAS signal is proportional to the total photoelectrically active surface of the particles, i.e. to d 2, while the total mass is proportional to the volume, i.e. to d 3. Plots of the ratio PAS signal to mass concentration vs. particle diameter exhibited clearly the expected lid dependence. Therefore the product of PAS signal and particle diameter and the concentration of the species should be correlated by a linear relationship, as shown for CoO in Fig. 6. A correlation analysis (Table 4) was done for the aerosol systems under study, and the high correlation coefficient confirmed the expected linear relation. With the results of Table 4 a rough estimation of detection limits for CoO, NiO and CdO is possible. The estimated detection limit for CdO was 800 ng/m 3, 250 ng/m 3 for NiO and 180 ng/m 3 for CoO. According to Eq. (7) a quantitative analysis of the aerosol is possible, if the particle size distribution is known.

A fast method for obtaining the particle size distribution is based on a screen-type diffusion battery with a condensation nucleus counter [ 3 4 - 38]. The extension of the PAS with an on line size classification of particles would allow an on line determination of the aerosol mass concentration. This aspect o f the PAS was studied with another simple experiment, in which a total of 30 monodisperse CoO aerosol samples with particle diameters between 2 0 - 7 0 nm were used for calibration and prediction. The PAS signal and the mass concentration of the samples were determined as described above. For a monodisperse aerosol Eq. (7) can be simplified to a linear regression model, i.e. for each wavelength the mass concentration was regressed on the product of PAS signal and particle diameter. For the calibration set 24 samples were used, while the remaining 6 samples were taken as independent validation samples. For each wavelength the regression model was applied to the prediction samples, yielding a prediction error R M S P (root mean square error of prediction) as shown in Table 5. Figure 7 illustrates the difference between the actual mass concentration and the predicted values. Applying multivariate regression techniques like M L R and PLS to the data set, all wavelengths can be used for calibration and prediction. In M L R a linear relationship is established between one or more dependent variables (here concentration of CoO) and some independent variables (PAS signal at wavelength 2k), which are measured for n objects (samples). The regression coefficients are estimated by the least squares method. In PLS the regression factors are linear combination of the variables, which are estimated by the principal components. These factors are estimated in such a way that they describe at the same time the variation that is relevant

734 Table 4. Analysis of the correlation between mass concentration and the product of PAS signal and particle diameter Compound

Diameter range ~ [nm]

Data points

CdO CoO

18-70 20-70

13 30

NiO

18--70

10

2i wavelength [nm] '185 185 214 229 185 214 229 254

rb

L~//L2 ~

0.993 0.995 0.995 0.994 0.987 0.984 0.984 0.975

0.998/0.977 0.998/0.990 0.997/0.989 0.997/0.987 0.997/0.946 0.996/0.933 0.996/0.932 0.994/0.894

Particle diameter was varied within this range b Correlation coefficient, significant at a level cq2) = 0.05 Upper (1) and lower (2) confidence interval, at a level eft2) = 0.95

Table 5. Prediction error for the determination of mass concentrations in monodisperse CoO aerosols

3.70 -3.30

Linear regression on single wavelength 185 nm 214 nm 229 nm Multivariate Calibration MLR PLS

RMSP" 43 ng/m 3 23 ng/m 3 24 ng/m 3

g

J

2.90

J 7 J

2.50 J c o

J

21o

i/ n

58 ng/m 3 31 ng/m 3

J

1.70

J

IE

J

1.30

a The RMSP (root mean square error of prediction) is defined as: rt k

i~ pI=I

l! )

Iv = Number of samples for prediction q = Actual mass concentration of sample i = Predicted mass concentration of sample i

for modelling the response matrix as well as the concentration matrix. In comparison to PLS multiple linear regression performed relatively poor, indicated by the R M S P values (Table 5). This is due to the fact that the same physical p h e n o m e n o n is responsible for the signal at each wavelength 2k. Therefore all wavelengths are highly collinear, which inevitably produces unstable M L R estimators. The collinearity is revealed by the correlation matrix of the CoO calibration d a t a (Table 6). As the intrinsic dimension o f the PAS signal vector is nearly one, the calibration d a t a matrix is modelled by the PLS algorithm with one factor. The one factor model was confirmed by cross validation and accounted for 94% o f the variation in the data. F o r both calibration methods, PLS and M L R , Fig. 8 shows the difference between the observed and the predicted CoO mass concentration. Although the simple linear regression model showed a superior prediction ability, in real applications PLS calibration should be preferred due to the noise reduction and the possibility of simple outlier detection. Owing to its conceptual appeal and c o m p u t a t i o n a l speed the application of PLS for quantitative analysis with the PAS sensor is a promising approach, which will be investigated further.

J

u

o.9o o_

•

• I

o.so ) ' / 0,50

,

I

1.30 0.90

,

I

~

I

1.70

*

I

Z90

2.10 2.50

Actual mass concentration [jag/m] ]

Fig. 7. Difference between actual and predicted CoO aerosol mass concentration for calibration by linear regression using only one wavelength 2k (at 185 n m , • 229 nm, [] 214 nm; particle diameter between 2 0 - 8 0 nm)

4.2 P a t t e r n recognition analysis

To explore the ability of the PAS sensor to differentiate between several photoelectrically active aerosol systems a d a t a set o f 80 objects was compiled, consisting o f m o n o disperse aerosol systems o f various diameters. Three highly photoelectrically active aerosols Cu20, N i O and CoO (Table 1) were chosen, while the fourth class o f objects consisted o f C o O / N i O mixtures with different ratios o f particle diameter and n u m b e r concentration. Within the experimental error the characteristic pattern vectors o f the Cu20, N i O and CoO aerosols are independent of particle diameter (Fig. 9). Given the linear additivity o f the PAS signal as described above, this will also hold for polydisperse aerosols and their characteristic response vectors. The pattern vector o f the C o O / N i O mixture is naturally only constant for an a p p r o x i m a t e l y constant ratio of the total photoelectrically active surface o f the components. As a graphical representation o f the d a t a matrix (80 x 4) is difficult to realize, a principal c o m p o n e n t analysis (PCA) was done o f a representative part o f the d a t a set, which was also used later as training set for the K N N pattern

735 3.50

Table 6. Correlation between the variables (wavelengths) in the CoO calibration data matrix"

3.00 Z50

Wavelength

185 nm

214 nm

229 nm

185 nm 214 nm 229 nm

1.00

0.998 1.00

0.997

ZOO 1.50 04

0.999 1.00

U_

" Correlation is significant at a level c%) = 0.01

3.70

1.OO 0.50 0.OO -0.50

m

-I.OO

c3 []

-1.50 _z_oo~5°

[]

,,,..,..' I

-ZOO

cJ~

+-

[]

I

-I.50

I --1.00

I --0.50

I 0.013

I 0.50

I 1.00

I 1.50

I 2.00

2.50

3.30

o~ *s

Fig. 10. Score plot of PCA (11 CoO, • NiO, [] Cu20, * NiO/CoO mixture; the origin is depicted by a +)

/ /

2.50

/

c

o

2.10

c~ E

1.70

7

/ /

k 1,30

c. O_

Factor 1

J

/

2.g0

/

- -

0.90

"<

0.5o 0.50

I

,

!

,

I

1

,

0.90

Acfuo[

I

,

,

2.10

1.30

I

,

I

2.90 2.50

1.70

mass c o n c e n t r a t i o n

3.30

[pg/m 31

Fig. 8. Difference between actual and predicted CoO areosol mass concentration for calibration by PLS ( • ) and MLR ([]) (particle diameter between 2 0 - 80 nm)

0.85

0.80 I 0.75

0.70 0.&5 _~ 0 . 6 0 ~n 0.55

~< o.5o gN o.45 0.4o E 0.3,5

0.25 O.%~°n.OO •

I 30.OO

]

40.OO

I 50.00

I

60.OO

I 70.OO

I

80.00

I

1OO.(30

90.00

PQrticl.e diameter Into]

Fig. 9. Relation between particle diameter and normalized pattern vectors of monodisperse CoO aerosols (11 185 nm, [] 214 nm, • 229 nm)

recognition method. P C A is a multivariate display technique [39-41], which aims at reproducing objects and variables in a low-dimensional c o m m o n factor space, so that a better interpretation o f similarities and dissimilarities between objects and/or variables is possible. In this case with two latent variables, which were calculated as described in [24], 96% of the variance in the data set was described, i.e. a simple 2-dimensional representation of the data is achieved. Both factors represent a sepa-

ration of the samples according to their characteristic photoelectrical activity at the four wavelengths of the PAS array. The projection of the objects, i.e. the pattern vectors, onto the latent variables is shown in a so-called score plot (Fig. 10). The CoO, NiO and C u 2 0 samples are clustered in three distinct groups, while the samples o f the CoO/NiO mixture are spread between the NiO and the CoO clusters due to their partial inhomogeneity in total photoelectrically active surface. The separate grouping of the objects in the factor space confirmed that the PAS signal contains sufficient information to differentiate between the characteristic response pattern of several aerosol systems. To obtain additional information about the assumed data structure a cluster analysis [42] of the complete data set was done. As no prior information about the data set is used, the clustering of the objects will confirm the hypothesized homogeneity and number of classes and detect possible outliers. Although the calculated clusters implicate a certain structure in the data matrix, it should be noted that a chemical significance of this structure is not automatically implied. The four main clusters are isolated by cutting the highest links as depicted by the dashed line in the dendogram (Fig. 11). The CoO, NiO and CuzO samples form three distinct clusters, apart from three C u 2 0 and NiO samples with extreme pattern vectors. The samples of the CoO/NiO mixture in which one of the components is dominant are linked either to the CoO or NiO cluster. The mixture samples with a constant ratio of total photoelectrically active surface form a cluster, which is finally linked to the NiO cluster due to signal of the mixture at 254 nm, which is characteristic for NiO (see Table 1). With the a priori information about the class membership of the data set (the training set) classification rules for supervised learning can be developed. Thus these rules are then applied to unknown samples (the test set). For simple pattern recognition the k-Nearest Neighbour method, a nonparametric, multiclass technique, was used. This approach assumes the existence of a local metric in the variable-space, which is related to the similarity between objects. A new object is classified according to its distance to the training set objects, i.e. the new object is assigned to the class of the majority of the k-Nearest Neighbours.

736 Table 7. Classification resuits for the k-Nearest Neighbour method

I

Class 1 (COO) Class 2 (NiO) Class 3 (CuzO) Class 4 (CoO/NiO) Total

Training set a

Test set

0/15 2/11 1/12 3/11 6/49 = 88% b

0/ 7 1/ 6 2/ 7 4/11 7/31 = 77% b

First values states the number of wrong classifications, while the second value is the total number of objects in the class a Training set was validated with the leave-one-out procedure b Total percentage of correct classifications in the data set

i m p a c t o r or applying a temperature gradient (as additional independent parameters) to the sensor inlet will possibly discriminate interferences.

5

i

1

I

1

I Z II

"

Fig. 11. Dendogram obtained from cluster analysis of the PAS data for pattern recognition (test and training set). 1 CoO samples; 2 CoO/NiO mixture samples with mainly CoO; 3 Cu20 samples; 4 NiO samples; 5 CoO/NiO mixture samples with mainly NiO; 6 CoO/NiO mixture samples with an approximately constant total photoelectrically active surface

The d a t a set was split into a training set consisting of 49 samples with k n o w n class memberships and a test set o f 31 samples for independent validation. Table 7 demonstrates that already with such a simple a p p r o a c h an excellent classification o f aerosol pattern vectors is possible.

4.3 Interferences The identification o f interferences is possible for compounds, which differ in their photoelectrical activity from the main c o m p o u n d at least in one wavelength. A c o m m o n interference like cigarette smoke, which is only p h o t o electrically active at 185 nm, can only be identified and corrected for, if the main c o m p o u n d is photoelectrically active at 185 n m and 214 nm. On the other h a n d an interference like diesel exhaust cannot be differentiated from a main c o m p o u n d like CoO, as both are photoelectrically active on three wavelengths. If the pattern vector of the emission source is nearly constant, the presence o f an interfering species can be deduced by a m a j o r change in the pattern vector. The limited resolution o f the P A S in this experimental configuration is caused by the fact that the intrinsic dimension of the signal vector is only one and therefore a deconvolution of multicomponent pattern vectors in mixtures is not possible. Using the sensor with an a p p r o p r i a t e

Conclusion

It was shown that the Photoelectric Aerosol Sensor is a valuable instrument for qualitative and quantitative analysis o f photoelectrically active polydisperse aerosols. The use o f a sensor a r r a y makes the differentiation between several aerosols already possible with simple pattern recognition methods. F u r t h e r m o r e emission sources, which contain several photoelectrically active species with an approximately constant ratio o f total photoelectrically active surface like the C o O / N i O mixture in our experiments, can be classified, too. W i t h knowledge of the distribution o f particle diameter the mass concentration o f photoelectrically active aerosols can be determined. Hence the extension o f the PAS with a diffusion battery and C N C is a promising approach, as it will allow simultaneously an on-line identification and quantification o f aerosol systems. W i t h the additional inform a t i o n a b o u t the particle size distribution mixtures o f different aerosols can be deconvoluted with standard techniques like factor analysis [43], as long as the particle size distribution o f the species are sufficiently different. This would improve the resolution o f the PAS, yielding a sensor for analysis o f m u l t i c o m p o n e n t mixtures of photoelectrically active aerosol systems.

Acknowledgements. The financial support of this study by the Deutsche Forschungsgemeinschaft and the Fonds der Chemischen Industrie is gratefully acknowledged. U.P. thanks the Studienstiftung des Deutschen Volkes for a fellowship.

References

1. Lain HF, Chen LC, Ainsworth D, Peoples S, Amdur MO (1988) Am Ind Hyg Assoc J 49:333 2. Hahn JU (1988) Staub-Reinhalt Luft 48:95 3. Glinsmann PW, Rosenthal FSS (1985) Am Ind Hyg Assoc J 46:391 4. Moreton J, FallaNAR (1980) Analytical sciences monographs, The Chemical Society. Burlington House, London, 52:17 5. Roeder M (1987) Metallverarbeitung 41:168 6. Eichhorn F, Oldenburg T (t988) DVS-Berichte 112: 25 7. Eichhorn F, Farwer A, Schwarzenbach E (1988) DVS-Berichte 112:29 8. Jockel W, Mistele J (1988) Staub-Reinhalt Luft 46:250

737 9. Gemeinschaftsarbeit aus dem BIA, St. Augustin (1986) StaubReinhalt Luft 44: M 6 6 - M98, Messen gesundheitsgef/ihrlicher Stoffe in der Luft am Arbeitsplatz 10. North MR, Haswell SJ (1988) J Anal Atom Spec 3:687 11. Ramamurthi M, Leong KH (1987) J Aerosol Sci 18:175 12. Kanapilly GM, Raabe OG, Newton GJ (1970) J Aerosol Sci 18:313 13. Ehrenhaft F (1910) Wien Ber 836 14. Twomey S (1975) J Comput Phys 18:188 15. Busigin A, van der Vooren AW, Philipps CR (1980) J Aerosol Sci 11:359 16. Markowski GR (1987) Aerosol Sci Technol 7:127 17. Kapadia AZ (1980) Data reduction techniques for aerosol size distribution measuring instruments, Ph.D. thesis, mechanical engineeringdepartment, University of Minnesota, Minneapolis, MN, USA 18. Niessner R (1986) J Aerosol Sci 17:705 19. Niessner R (1988) Fresenius Z Anal Chem 329:406 20. Niessner R, Wilbring P (1989) Anal Chem 61:708 21. Niessner R, Walendzik G (1989) Fresenius Z Anal Chem 333 : 129 22. Sharaf MA, Illman DL, Kowalski BR (1986) Chemometrics, Wiley, NY 23. Massart DL, Vandeginst BGM, Deming SN, Michotte Y, Kaufman L (1988) Chemometrics: a textbook. Elsevier, Amsterdam 24. Wold H (1966) Research papers in statistics. David F (ed) Wiley, NY, p 411 25. Wold H (1966) Multivariate analysis. Krishnaiah P (ed) Academic Press, NY, p 391 26. Wold S, Ruhe A, Wold H, Dunn WJ (1984) SIAM J Sci Stat Comput 5:735 27. Lorber A, Wangen LE, Kowalski BR (1987) J Chemo 1 : 19

28. Martens H, Karstang T, Naes T (1987) J Chemo 1:201 29. Martens H (1985) Multivariate calibration: quantitative interpretation of non-selective chemical data, Dr. Tech. thesis, Norwegian Institute of Technology, University of Trondheim, Norway 30. Kowalski BR, Bender CF (1972) Anal Chem 44:1405 31. Cover TM, Hart PE (1967) IEEE Trans Info Theory 13:21 32. Masuda S, Mizuono A, Tanaka S (1983) Proceedings 1st Symp on Aerosol Sci Technol Japan, p 35 33. Burtscher A, Schmidt-Ott A (1984) Sci Total Environ 36:233 34. Sinclair D, Hoopes GS (1975) Am Ind Hyg Assoc J 36:39 35. Sinclair D, Countness RJ, Liu BYH, Pui DYH (1979) Aerosol measurements. Lundgren DH et al. (ed) University Press of Florida, Cainesville, p 544 36. Cheng YS, Yeh HC (1980) J Aerosol Sci 11:313 37. Cheng YS, Keating JA, Kanapilly GM (1980) J Aerosol Sci 11:549 38. Wu JJ, Cooper DC, Miller RJ (1989) Rev Sci Instrum 60:258 39. Kvalheim OM (1987) Methods for the interpretation of multivariate data: examples from petroleum geochemistry, Ph. D. thesis, Department of Chemistry, University of Bergen, Norway 40. Wold S et al. (1984) Chemometrics - mathematics and statistics, NATO, ASI Series C, D. Reidel Publ. Comp., Dordrecht, 138:17 41. Wold S, Esbensen K, Geladi P (1987) Chemometrics and Intell Lab Syst 2:37 42. Massart DL, Kaufman L (1983) The interpretation of analytical chemical data by the use of cluster analysis, Wiley, NY 43. Malinowski ER, Howery DG (1980) Factor analysis in chemistry, Wiley, NY Received April 14, 1989