Image processing nanoparticle size measurement for determination of density values to correct the ELPI measures

Available online at www.sciencedirect.com

Precision Engineering 32 (2008) 88–99

Image processing nanoparticle size measurement for determination of density values to correct the ELPI measures Nicolas Coudray a , Alain Dieterlen a,∗ , Lo¨ıc Vidal b , Estelle Roth c , Gw´ena¨elle Trouv´e c , Sophie Bistac b a b

Laboratoire MIPS Group Lab.El, I.U.T. de Mulhouse, 61, rue Albert Camus, 68093 Mulhouse Cedex, France Institut de Chimie des Surfaces et Interfaces, 15, rue Jean Starcky, BP 2488, 68057 Mulhouse Cedex, France c Laboratoire de Gestion des Risques et Environnement, 25 rue de Chemnitz, 68200 Mulhouse, France Received 21 February 2006; received in revised form 12 March 2007; accepted 17 April 2007 Available online 5 July 2007

Abstract The ELPI is an electrical low-pressure impactor that classifies aerosol particles according to their aerodynamic diameter, and generates the amount of particles impacted on each of the 12 stages. This number depends on the measured current induced by the pre-charged particles and on the density which is given by the user and may not be a priori known. In addition, the density used by the software is considered to be similar for all stages. In this paper, a method to evaluate the density of the particles on each stage is proposed in order to consequently increase the accuracy of the results given by the software. The data needed are the aerodynamic diameter, the equivalent diameter and information on the form. The aerodynamic diameter is a range defined by the cut-off aerodynamic diameter of the stages of the ELPI. To measure the equivalent diameter and evaluate the form, an adapted procedure that used microscopy and image processing tools was set up with the study of two different polydispersed aerosols, silica and fly ash particles from wood combustion. This method was validated with Silica particles (ρ = 2.5 g cm−3 with the pycnometer): the density was found to be 2.2 g cm−3 and 2.4 g cm−3 for stages 2 (dae around 76 nm) and 3 (dae around 127 nm), respectively. The results match reality for fly ashes from wood combustion as well: ρ = 1.0 g cm−3 for the stage 2 and 1.9 for stage 5. © 2007 Elsevier Inc. All rights reserved. Keywords: Image processing; ELPI; Cascade Impactors; Nanoparticles; Microscopy; Density measurement

1. Introduction Fine particles are likely to pose a risk to environment and health since they can travel deeply into the respiratory apparatus. In this context, cascade impactors are of great interest since they can collect particles and measure the size distribution. They can be used in many fields, e.g. particles emitted from diesel engines, ambient aerosols or other aerosols from combustion sources [1–3]. In this study, the Electrical Low Pressure Impactor (ELPI) manufactured by Dekati Ltd., Tampere, Finland, collects particles from 29 nm to 10 ␮m, and classifies them into 12 size fractions. The ELPI can be divided into three parts: the particles are first electrically charged, then impacted on different stages according to their inertia, and finally, the number of particles is ∗

Corresponding author. Tel.: +33 3 89 33 76 61; fax: +33 3 89 33 76 05. E-mail address: [email protected] (A. Dieterlen).

0141-6359/$ – see front matter © 2007 Elsevier Inc. All rights reserved. doi:10.1016/j.precisioneng.2007.04.009

measured. The number of particles depends on the induced current measured at each stage, and it also depends on the density ρ, which has to be given by the user. For the calculation of the size distribution, the same ρ is used for each stage. However, in practice the density may vary from stage to stage. Particularly combustion processes produce aerosols with particles of different nature, each type having a specific density. Indeed, the mineral parts of these aerosols present densities from 2.0 to 3.5. They can reach 12 when the fuel contains a great heavy metal part. On the contrary, the flying ash density ranges from 0.6 to 1.2 and the soot one around 0.8. The choice of the density has a great influence since it may cause a fluctuation of more than 100% in the evaluation of the number of particles. The shape of the distribution is not modified but the number is affected. As a consequence for an accurate number determination, the density has to be set very precisely. The density ρ cannot be easily measured with classical experimental tools (like pycnometer) since there is not enough matter

N. Coudray et al. / Precision Engineering 32 (2008) 88–99

collected on each stage. This information is not only essential for the evaluation of the number size distribution done by the ELPI, but also in the characterization of the studied aerosol. This paper presents a method for the evaluation of ρ using imaging tools, as the density is proportional to the Stokes diameter and the aerodynamic diameter dae . dae is known from the ELPI since each stage is characterized by a lower cut-off aerodynamic diameter. As for the Stokes diameter, it is supposed to be the projected equivalent diameter deq for spherical particles (thus, in this paper, only deq is always used). The deq can be measured on images acquired with a scanning electron microscope whose resolution allows precise measure of the sizes. The repartition of the particles, the homogeneity and the regions of interest were selected according to observations done with a light microscope. First, ELPI operations which are essential to understand the benefit of our method are described. ELPI collects particles on substrate whose choice is directly connected to the analysis that needs to be done. In order to study light microscope images, coverglasses substrates were placed on each ELPI plates. Images from light microscope were taken and used to validate the effectiveness of this substrate and for general observations. Furthermore, this analysis made us aware of the losses that can occur in such an impactor. Having a valid support and protocol, and knowing the different losses, it became then possible to apply adapted image processing tools to extract the equivalent diameter. The protocol to extract the diameter and calculate the density of the particles is fully described in this paper. Validation of this procedure was tested with Silica particles. Then, improvement with loss estimation was studied and the method was extended to fly ashes density measurement. 2. Experimental 2.1. Principle of the electrical low pressure impactor The ELPI is made of three parts: the particles are firstly electrically charged by the corona charger, then, they are impacted according to their inertia in the impactor and finally, the size dis-

89

tribution is evaluated from the induced current of the pre-charged particles. 2.1.1. Corona charger The aerosol is sampled at a flow rate of 9.82 l min−1 and flies across the corona charger, whose higher part is made of a 5 mm long tungsten electrode which supplies a 5 kV positive voltage. The particles coming perpendicularly to the electrical field are charged with positive ions. The performance of the charger is characterized by the product between the penetration through the charger P and the average number of charges per particle n. This product Pn is used by the ELPI software and has been experimentally determined by the method described by Marjam¨aki et al. [4]. In our case, the values are given by Eq. (1): ⎧ 1.637 ⎪ ⎨ 5.941 × deq ; deq < 0.0239 ␮m 1.3201 ; 0.0239 ␮m ≤ deq <10 ␮m (1) P × n = 1.819 × deq ⎪ ⎩ 0.5909 79.732 × deq ; 10 ␮m ≤ deq deq being the particle diameter, which is the equivalent diameter of a spherical particle. This equation is then used to convert the current into size distribution. 2.1.2. Impactor A multi-stage cascade impactor (shown in the Fig. 1) is then used to classify the particles. A single stage is made of two collinear plates, the jet plate and the collection plate. The flow is going by the nozzles of the jet plate and turns sharply. Bigger particles that cannot follow the stream because of their inertia are impacted on the second plate where the substrate has been fixed. Smaller particles that follow the stream are impacted on the next lower stages. The size and the number of the nozzles vary from plate to plate. It permits the impaction of particles of different inertia and governs the size distribution. The repartition of the particles on the substrate is directly related to the number and lay out of the nozzles.

Fig. 1. Photography and Scheme of the three lower stages of the ELPI with induced current of impacted (Iinertie i ) and lost particles (Ii/k ).

90

N. Coudray et al. / Precision Engineering 32 (2008) 88–99

Table 1 Lower aerodynamic and equivalent cut-off diameters of the ELPI stages for different densities

ρ = 2.1 ρ = 2.5 ρ = 2.9

Stage

1

2

3

4

5

6

7

8

9

10

11

12

dae (␮m) deq (␮m) deq (␮m) deq (␮m)

0.029 0.014 0.012 0.011

0.057 0.029 0.025 0.022

0.095 0.051 0.044 0.038

0.158 0.090 0.079 0.069

0.264 0.161 0.142 0.127

0.384 0.243 0.217 0.196

0.616 0.402 0.363 0.332

0.953 0.634 0.575 0.529

1.610 1.087 0.990 0.913

2.400 1.631 1.489 1.377

4.010 2.742 2.506 2.321

9.960 6.848 6.269 5.815

Each stage is then characterized by a specific aerodynamic diameter: it is the diameter of a spherical particle of unit density that has the same settling velocity of the considered particle (Eq. (2)).  ρC(deq ) dae = deq (2) ρ1 C(dae ) with ρ the particle density, ρ1 the unit density (1 g cm−3 ), C(dx ) is the Cunningham factor, deq the Stokes diameter (which is the equivalent diameter of the spherical particle). The impactor covers sizes from about 29 nm to 10 ␮m. Thus, according to their inertia, the particles are collected or not on each plate. Each stage is characterized by the cut-off diameter dae50% , that is the cut size with 50% efficiency. Characteristics of ELPI stages are given in the Table 1 (first two line). 2.1.3. Output The charged particles induce a current when passing by a stage. According to the electromagnetic laws, the current induced is positive for the particles coming and negative for the particles leaving the stage. Thus, the measured induced current for a given stage is null if the upstream flow equals the downstream flow. If not, it means that some particles have impacted and the non-null induced current measured is proportional to the number of particles collected. Each stage is electrically insulated with Teflon and a real-time measurement is obtained by electrometers. This current is then converted into a size distribution of the aerosol. The relation between the concentration Ci of particles per stage i and the current I is given by the Eq. (3). Ci =

I PneQ

(3)

where Ci is the number concentration of particles per second of sampling (cm−3 s−1 ), P the penetration through the charger, n the number of charges per particle, e is the charge of an electron (1.602 × 10−19 C) and Q is the flow rate (9.82 l min−1 ). C is directly proportional to the number of particles collected N (Eq. (4)). N = Ci tQ

(4)

where t is the sampling time (s). The factor Pn characterizes the charger efficiency and it depends on deq which is calculated by Eq. (5). deq = 

dae ρC(deq )/ρ1 C(dae )

(5)

For a given stage, dae is the geometric mean of the lower and upper cut off diameter. For instance, for stage 5: √ (6) dae = 264 × 384 = 318 nm ρ is needed to solve deq and has to be given by the user. 2.1.4. Challenge: density evaluation Thus, the first problem is to evaluate the density ρ. The density cannot be easily measured with classical experimental tools (like pycnometer) since there is not enough matter collected on each stage. This information is not only essential for the evaluation of the number size distribution done by the ELPI, but also in the characterization of the studied aerosol. The second is that this information is considered by the software to be identical for each stage. The assumption of constant density over the whole range of diameter is known to be a limitation of the instrument [5]. Indeed, ρ may vary according to the sampled aerosol and its chemical compounds suggesting that ρ can vary for each stage. The imprecise and global knowledge of the density ρ will induce errors in the evaluation of deq , thus in the characterization of the aerosol. The proposition of this study is to measure deq for each stage i to deduce ρ for each stage. The equivalent diameter of the particles impacted on the substrate was measured with image processing techniques. 2.2. Correction process The goal of the control process is to enhance the evaluation of the particle size distribution analysis by evaluating ρ. Because usual tools used by chemists are not appropriate to measure very low volumes of matter, image processing can achieve the determination of ρ by measuring the equivalent diameter of particles impacted by the ELPI. For chemical analysis, most of the time, quartz fibres or Teflon filters are generally used to collect matter on each ELPI stage. These substrates which are not transparent do not permit the acquisition of images. A new suitable substrate in glass was used to collect the particles and permit image acquisition with optical and electron microscopes. A first part of the study was to evaluate the substrate capacity to collect particles (paragraph 3.1). The process based on image analysis is described on Fig. 2: Optical Microscopy (OM) gives global information about the repartition of the collected particles for each stage, while the higher resolution of the Scanning Electron Microscope (SEM) permits to extract the size of the particles. This measure is combined with the aerodynamic diameter of the studied stage and Eq.

N. Coudray et al. / Precision Engineering 32 (2008) 88–99

Fig. 2. Control process of the ELPI. Regions of interest are identified with OM before the acquisition with the SEM. Measure of deq is obtained with image processing tools on particles observed on SEM images to calculate the density ρ. This value can be reintroduced into the ELPI software to correct the size distribution.

(2) to calculate the density. This new density value introduced in the ELPI macro adjusts the size distribution. In the next paragraph, the image processing protocol developed to extract the required information is described. 2.3. Image processing tools The FEI model Quanta 400 Scanning Electron Microscope was used to measure deq related to its nanometer resolution. The

91

Fig. 3. Example of piled up silica particles on stage 2, at a magnification of 100,000×.

images were challenges that we had to take with ImageJ image processing toolbox (http://rsb.info.nih.gov/ij/): • First, the complexity of the pictures: in most cases, the particles are piled up and so close that it may be difficult to differentiate the border of each particles (Fig. 3). • Secondly, the diversity of the images: some parameters vary from stage to stage (magnification, amount of particles more or less close on Figs. 3–5).

Fig. 4. Automated processing of separated particles of Silica on stage 8, magnification 5000×; example 1: initial picture (1); picture after segmentation (2); picture after watershed algorithm (3); labelization of the measured particles (4).

92

N. Coudray et al. / Precision Engineering 32 (2008) 88–99

Fig. 5. Automated processing of separated particles of Silica on stage 2, magnification 100,000×; example 2: initial picture (1); picture after segmentation (2); picture after watershed algorithm (3); labelization of the measured particles (4).

These characteristics led to setting up adapted processes to treat two kinds of images: images where the particles are separated and those where the particles are piled up. Once particles are clearly discriminated, adequate measures and selection were made. 2.3.1. Processing of separated particles An automated process made up of four steps was set up to measure the diameters: pre-treatment, segmentation, posttreatment and measures. The goal of the first step (pre-treatment) is to reduce the noise that can be added when the images are acquired. Median filter was chosen since it is a good compromise between noise reduction and gradient alteration. For the segmentation, the thresholding method was used since the particles are not amassed and background is constant. Besides, the selection of the threshold values is easy since the grey level of the objects and the grey level of the background are different enough. However, some particles are very close and may be considered as one object after the segmentation. Thus, the watershed algorithm was used as a post-treatment to tackle these cases. Next, the measures were only done on particles made with at least 50 pixels to ensure good precision of the measures and get rid of the possible isolated pixel (noise remaining). Fig. 4 presents a first example of image (stage 8, magnification 5000×) where this process can be applied. Labels were placed on particles with more than 50 pixels, thus 44 particles.

Others were rejected since they are either from noise or particles too small and they may induce a too important error. Eight out of 44 particles can be considered as non spherical (see Section 2.3.3. for the measure description). The 36 particles remaining have a mean diameter of 20.7 pixels (1106 nm) with a standard deviation of 5.3 pixels (283 nm). Fig. 5 shows another example with another stage (2) and another magnification (100,000×). Rejecting the particles with less than 50 pixels, 213 particles were measured, 174 of them are considered to have low circularity. The 39 remaining particles have a mean diameter of 12.6 pixels (34 nm) with a standard deviation of 3.4 pixels (9 nm). The two examples show that the methods give credible results for different magnifications (5000 and 100,000) and different stages (P8 and P2, respectively). Furthermore, the method is robust, pretty fast and allows many measures at a time (about 35 measures per image in these cases). However, the process is not adapted to regions where particles are amassed, as shown on Fig. 6. Another process was applied for piled up particles. 2.3.2. Segmentation of piled up particles The main problem for the piled up particles is the segmentation phase. The aim of the segmentation is to discriminate the objects that had to be measured. However, when particles are pilled up, we have to select the particles one after the other man-

N. Coudray et al. / Precision Engineering 32 (2008) 88–99

93

Considering aliasing phenomena, the precision of the measure was evaluated to be 6% for deq . It leads to a 12% error on the estimate ρ. The form factor informs on how spherical an object is: FF = Fig. 6. Pilled up particles (left) and result of the automated processing (right).

ually or with the snake method. The snake does not require any a priori knowledge of the object except its position by drawing a ROI (Region Of Interest). The selection is modified dynamically to minimize the energy function:

  i i i Esnake = γ(i)Eimage (7) [α(i)Econt + β(i)Ecurv ]+ i i Econt

i i Ecurv

where and are internal energy terms that force the i contour to be continuous and smooth. Eimage is known to be the external energy that attracts the contour to the edges. It is hence gradient dependant and the contrast enhancement previously done improves the performance of the snake. The terms α, β and γ regulate the weight of the corresponding energy terms. This semi-automated procedure was based on the tool developed by Boudier [6] for ImageJ (example Fig. 7), in which gradient threshold, regularization and number of iterations are parameters that must be chosen and are image dependant. The gradient threshold parameter can be chosen more easily if the contrast of the image is previously enhanced. 2.3.3. Measures Once the objects discriminated, the measures are possible. The equivalent diameter is defined as the diameter of a sphere having the same volume than the considered particle. However, the opacity of particles does not allow 3D acquisition. As a result, it was restricted to 2D-measurement which may lead to extra error. In 2D cases, the equivalent diameter is to be a circle of equal area of the projected area A of the particle seen on the image:  Area deq = 2 (8) π

Fig. 7. Result of the snake method selection based on a rough initial manual selection.

4πArea Perimeter2

(9)

If FF = 1, then the particle is spherical. The lower the FF, the more elongated the particle. If the 2D particle is spherical, then, the 3D particle is likely to be spherical as well and the deq measured on the 2D images can be considered as the deq of the 3D particles. If the particle is not spherical, then the deq is not considered for the calculation of ρ. Furthermore, segmentation, aliasing and the method to measure the parameters (perimeter, area) induce an imprecise evaluation of FF. Particles were experimentally considered to be spherical if FF > 0.8. Other particles are not kept for the calculation of the density. Allowing this procedure, for a given plate, the area of the spherical particles is measured and the equivalent diameter is deduced. Then, a size distribution of the particles impacted on the stage considered is obtained. The distribution is then fitted to a Gaussian from which the mean value is extracted. The same curve is supposed for the aerodynamic diameter distribution and the average value for dae is extracted from this curve as well as the cut-off diameters. 3. Results Before SEM image processing, OM image processing is used to evaluate the new impaction substrate and to gain a better understanding of the impaction process (Fig. 2). Then, the method is validated with Silica particles of known density. This analysis reveals the fine particles losses that may occur and influence the observations and, consequently, the measures. In the Section 3.3, a fine particles loss model is therefore proposed to choose the particles that need to be taken into account for the 3D equivalent diameter measurement of the particles impacted on the considered stage. The whole process is then applied to fly ash particles from wood combustion. 3.1. Glass substrate evaluation The collection of the particles was made with substrates placed on the collection plates. It has been proved that rebounds of particles are minimized if the substrate has cavity or fibers [7] and that roughness and porosity improve the collection efficiency [8]. However, Teflon, quartz or aluminum substrates usually employed are not suitable for the present study, since they are either opaque or fibered (thus heterogeneous). In addition, it was observed that fibers modify the aspect of the particles. Condensation phenomena of mineral particles (salts) around the fibers occurred. These problems prevent any reliable measurement of the characteristics of the images. Coverglasses of 22 mm diameter round purchased from ROTH were fixed by capillarity with 10 ␮l water on each ELPI plate. The measurements done by the ELPI did not appear to be biased by this method of collection according to the exper-

94

N. Coudray et al. / Precision Engineering 32 (2008) 88–99

Fig. 8. One spot optical microscope observation of particles impacted on stage 3 (left) and 8 (right).

imented users, the currents measured and the shape of the distributions. To study the geographical repartition of the particles, a BX5Olympus light microscope, allowing a non destructive observation of the substrates, was used. The microscope is equipped with a 10× objective and a CCD camera. The observations are very different: depending on the aerosol and the stage, the particles are more or less numerous and more or less close to each other (Fig. 8). The repartition can be considered at two levels: the spot level and the whole stage level. After passing through the nozzles of the jet plate, bigger particles impact on the collection plate which the glass substrate was fixed on. Under each nozzle, there is a spot of particles (Figs. 8 left and 9). The shape of the spots may vary according to the substrate and according to the jet plate. The form and size of the spot help the evaluation of the efficiency of a substrate [9–10]. However in previous papers, the global view of a stage of the ELPI was not yet discussed. On stage 3, several connected images are combined to obtain a broad field of view containing several spots. Fig. 9: elliptical spots of particles just laid below the nozzles of the jet plate are observed. Between these spots, grey regions show that some particles were collected outside the expected regions. These particles are geometrically located between the spots, exactly where the streams from the different nozzles may meet. The most probable hypothesis is that turbulences are created by the different streams as explained in Fig. 10. Particles collected because of the turbulences are prob-

Fig. 9. Mosaic (7.2 × 2.5 mm) of 40 images of stage 3 observed by light microscopy.

Fig. 10. Scheme of stream and particles deposition by turbulences.

ably smaller particles not impacted that follow the stream. The size of these particles were measured for stage 3 and compared to the size of the particles in the spot and to the size of the particles impacted on the next stage (stage 2). The results shown on Fig. 11 corroborate the hypothesis that the size of the particles collected out of the spots is smaller than the size of those in the spots. However, for the upper stages, the particles concentration in the stream is higher and the size distribution of the incoming particles is more widespread. Consequently, the size distribution of the particles collected because of the turbulences is expected to be more widespread as well. These observations lead to several comments: • First, the size distribution of the particles impacted out of the spot is different from the size distribution of the particles inside the spots.

Fig. 11. Number size distribution of particles in the spots of stage 2, 3 and out of stage 3.

N. Coudray et al. / Precision Engineering 32 (2008) 88–99

95

• Secondly, these particles are probably not impacted because of their inertia but collected because of the turbulences. Thus, their aerodynamic diameter is not characteristic of the dae of the stage. • Thirdly, according to the images (Fig. 9) these particles seem to be pretty numerous, which probably significantly biased the current measured by the ELPI. This analysis confirms the effectiveness of the glass substrate. The observation of the turbulence phenomena and the size of the particles located out of the spots led us to consider the other sources of uncertainty. Therefore, only particles located in the spot were measured to evaluate ρ. 3.2. Validation with Silica particles The method was firstly applied on polydispersed particles of Silica purchased from SOVITEC (ref 050-5-215) whose density was measured with a pycnometer at 2.5 g cm−3 ± 0.4 (thus ± 15%). According to the precision of the measure, the density is expected to stand between 2.1 and 2.9 g cm−3 . Considering Eq. (2), the expected values for the equivalent diameter deq can be deduced (Table 1), with dae being the lower cut-off aerodynamic diameter of the stages (according to the characteristics of the ELPI). For instance, for stage 2, deq should vary from 22 nm up to 51 nm, for the minimum and maximum values of the cut-off diameters. However, since the cut-off diameter is given for 50% efficiency, some particles collected are bound to be slightly smaller and bigger than these values. Two experiences were done: in the first, smaller Silica particles were collected (up to stage 4; Fig. 12); in the second, bigger Silica particles (from stage 4; Fig. 13). For the experience 1, a mass of 5 mg of silica was placed in a tubular quartz pan which was introduced in a tubular furnace heated at 600 ◦ C. A total air flow rate of 315 N l h−1 was passed through the furnace. The electrical impactor was connected at the end of the furnace and collected the silica aerosol during 30 min. With ρ = 2.5 g cm−3 , the number distribution given by the ELPI is represented on Figs. 12 and 13 where each bin represents a stage. From left to right: filter stage, stage 1, stage 2,. . .,stage 10, stage 11. On experience 1, the particles are mainly collected

Fig. 12. Number size distribution of Silica—Experience 1.

Fig. 13. Number size distribution of Silica—Experience 2.

on stages 1, 2 and 3 (Fig. 12). However, only the particles on the stages 2 and 3 were studied since the resolution of the Scanning Electron Microscope (SEM) is not sufficient for stage 1 (deq expected to be less than 25 nm). In a primary approach, the analysis of the size distribution of the 1098 measures made on stage 2 is represented on Fig. 14. The three more important bins represent the particles which size is exactly within the expected range (from 25 nm to 44 nm, with ρ = 2.5 g cm−3 , Table 1), that is 74% of the measured particles. Considering a density variation of 15%, then 86% of the particles are within the expected range (22 nm to 51 nm, Table 1). In a second approach, the description of the distribution is used to determine the mean density ρ of the stage 2. The repartition of the particles can be considered as Gaussian with a mean value at 37.9 nm and a 99% correlation. Thus, the average equivalent diameter of the particles impacted on stage 2 (deq = 38 nm) was used to compute the density according to Eq. (2). In this equation, dae for stage 2 is an evaluation of the mean cut-off aerodynamic diameters: dae =

dc2 + dc3 57 + 95 = = 76 nm 2 2

(10)

dc2 being the lower cut-off diameter of stage 2 and dc3 the lower cut-off diameter of stage 3 (Table 1). Thus, the value of the average density of the particles is 2.2 g cm−3 . This value stands in the range defined by the pycnometer results.

Fig. 14. Number size distribution of Silica particles impacted on stage 2 (1098 measures).

96

N. Coudray et al. / Precision Engineering 32 (2008) 88–99

thus the result cannot be taken into account. The fact that these distributions are so wide probably means that stages 6 and 7 are affected by the fine particle losses and by the rebounds. Stages 8 to 10 are probably affected by the losses too, but these particles were too small compared to the impacted ones to be measured. Then, according to this test done with Silica particles, the method used to determine the density of the particles is valid when the correlation factor is significant (94%). 3.3. Analysis of the losses (versus inertial impaction) Fig. 15. Number size distribution of Silica particles impacted on stage 3 (79 measures).

The same reasoning is applied to the particles measured on stage 3. Fig. 15 shows the repartition of the 79 measured particles, where 94% of the particles stand in the expected range, from 38 nm to 90 nm. As for the equivalent diameter of this Gaussian distribution, it was identified to be 62.4 nm, with a correlation factor of 94%. Since the cut-off aerodynamic diameters of stage 3 are 95 nm and 158 nm (mean dae = 127 nm), the density calculated with Eq. (2) is ρ = 2.4 g cm−3 . This value confirms the potential of the method. The second experience done with Silica permitted an analysis of the distribution of bigger particles (Fig. 13). A 250 mL closed flask containing a silica sample of 30 mg was placed on line the impactor. The flask was shacked at room temperature during few minutes and the silica aerosol was pumped by the ELPI. The images of the particles impacted on the upper stages were acquired with a SEM. The size of the particles observed on stages 6–10 varies significantly, from nanometer particles up to micrometer particles. This corroborates the fine particle loss phenomena [12]. The proposed method is used to extract deq and evaluate ρ for the particles of the upper stages. All the results are summarized in Table 2. The first two columns concern the experience 1 (stages 2 and 3), and the last five the experience 2. The density values which correspond to the ones expected are in bold. For the stage 9, the distribution is more lognormal, shifted to the lower sizes, while the dae considered is still the mean value of a Gaussian distribution. This can explained while the density is slightly greater than the maximum expected (2.9 g cm−3 ). Moreover, the correlation factor is slightly lower than the one of the ‘good’ results. As for the stages 6 and 7, the distribution is more spread out and the identification with a Gaussian is not valid anymore. The correlation factor is indeed quite low,

As it was shown in paragraph 3.1, the impaction is not perfect and some particles observed on a given stage may not have been impacted by inertia. In this paragraph, the main phenomena recorded in the literature are studied: we consider the particles that are bigger than those expected and those smaller (fine particles). First, the particles may bounce and then be collected by a lower stage. Secondly, particles already collected can be re-entrained in the stream by coming particles [5]. Finally, agglomerates of particles appear too. This phenomenon is well known in aerosol science [11]. Theses three incidents imply the collection of bigger particles whose size is not representative of the aerodynamic diameter of the considered stage. The nonideal collection of particles on the lower stages is mainly due to these problems. In our cases, it has been visually quite easy to differentiate the bigger particles and those impacted by inertia (Fig. 16). Fine particle losses were previously described by Virtanen et al. [12] when aerosols were collected by the ELPI. These are caused by charge effect or diffusion (Brownian motion). They vary from stage to stage and it is important to discriminate the particles lost and those impacted, since only the size of the impacted particles is significant of the aerodynamic diameter. The ELPI software gives the measured original current and a corrected current: an example is displayed on Fig. 17 for an experiment done on fly ash aerosol from wood combustion [5]. The corrected current is the current that should have been measured if there was no fine particle loss. Here, a method is proposed to identify the quantity of particles impacted by inertia and the fine particles lost.

Table 2 Results of the analysis on Silica particles Stage

2

3

6

7

8

9

10

dae (nm) deq (nm) Correlation ρ (g cm−3 )

76 38 99% 2.2

127 62.4 94% 2.4

500 800 85% 0.4

785 858 81% 0.8

1282 834 97% 2.2

2005 1118 93% 3.1

3205 1771 95% 2.2

Stages 2 and 3 for experience 1 and stages 6–10 for experience 2; density values within the expected range are in bold.

Fig. 16. MEB observation of a region of a spot on stage 2—Small particles (impacted) and, on the left hand-side, three bigger particles (bounce, re-entrainment or agglomerate).

N. Coudray et al. / Precision Engineering 32 (2008) 88–99

97

In a first approach, it is supposed that the particles are equally lost on the eleven upper stages, then: I1/i =

Icorrected

− Iinertie 11

1

1

Icorrected 1 − I1 , 11

=

∀i > 1 (18)

Related to Eqs. (14) and (18), the percentage of losses at stage 2 can be numerically calculated: %Losses 2 = 100 ×

Icorrected

1

− I1 /11

(19)

I2

Similarly for this stage 2, the corrected current is expressed as: Fig. 17. Measured I and corrected current Icorrected for log normal number size distribution, each bin corresponding to a stage. Fly Ash aerosol.

Icorrected

2

= Iinertie

2+

12  i=3

The two distributions of the currents (Fig. 17) are the base of this method: Ii is noted for the measured current at stage i and Icorrected i for the corrected current at the stage i. The reasoning is also based on the truncated representation of the ELPI as shown on Fig. 1 and will be developed for the three lower stages. The aerosol is introduced from the upper stages and it flows down the impactor. The bigger particles (Fig. 1) impact on stage 3 and induce a current Iinertie 3 . The particles that should have impacted on stage 2, but have been lost on stage 3 induced a current I2/3 on stage 3 (medium particles on Fig. 1, stage 3). Those who should have impacted on stage 1 but have been lost on stage 3 induced a current I1/3 on stage 3 (smaller particles on Fig. 1, stage 3). Thus, the current measured originally on stage 3 is:

combining with Eq. (18):

I3 = Iinertie

with (from Eq. (18))

3

+ I2/3 + I1/3

(11)

Thus, the percentage of the losses compared to the measured current is: I2/3 + I1/3 %Losses 3 = 100 × Iinertie 3 + I2/3 + I1/3 I2/3 + I1/3 = 100 × I3 I2 = Iinertie

2

+ I1/2

%Losses 2 = 100 ×

(13) I1/2 I1/2 = 100 × Iinertie 2 + I1/2 I2

(14)

and for stage 1: I1 = Iinertie

(15)

1

%Losses 1 = 0%

(16)

The corrected current is the current that should have been measured, without the losses. Thus, for stage 1, it is the particles impacted on stage 1 and those lost on stage 2–12 (if there are 12 stages) that should have reached stage 1: Icorrected

1

= Iinertie

1

+

12  i=2

I1/i

which leads to Icorrected I2/i = =

I2/i =

Icorrected

Icorrected

(17)

− Iinertie 2 10 2 − (I2 − I1/2 ) , 10 2

2

− (I2 − Icorrected 10

∀i > 2

1

− I1 /11)

(21)

,

∀i > 2 (22)

Thus, the percentage of losses is all known for stage 3: %Losses 3 = 100 ×

I2/3 + I1/3 I3

(23)

Icorrected 1 − I1 11 and (from Eq. (22)) I1/3 =

I2/3 =

Icorrected

(12)

Similarly for stage 2, we found:

(20)

I2/i

2

− (I2 − Icorrected 10

(24)

1

− I1 /11)

(25)

The same method is then applied to all the stages i to evaluate %Losses i. In this study, the number N of particles is a more relevant information than the current I. The previous currents calculated are converted into number using the Eq. (26). N=

It Pne

(26)

where N is the number of particles, I the current (A), where t is the sampling time (s), P the penetration through the charger, n the number of charges per particle, e is the charge of an electron (1.602 × 10−19 C). Related to Eqs. (1) and (2), the transformation from I to N is dependant on the density ρ. This transformation from I to N is applied to each equation to find the number of particles impacted and the number of particles lost for each stage. Based on the measures made on an experience with fly ashes, Fig. 18 represents the importance of the losses %Losses i for each stage. As it was expected, the fine particle losses are much more important on upper stages. The way the measured current

98

N. Coudray et al. / Precision Engineering 32 (2008) 88–99

Fig. 18. Percentage of number of lost particles at each stage and influence of the density.

is biased highly depends on the experience. From stage 5 up to 11, more than half of the particles are lost particles in this case. In this example, the curves do not seem to be very influenced by the density of the particles or the experience (Fig. 18). Related to this study, it is possible to identify which particles have to be considered for each stage for the measure of deq and consequently the measure of ρ. This approach is helpful to establish which particles have to be analyzed by the image processing tools. For the lower stages, no correction needs to be done. For the intermediate stages and upper stages, the losses are important and a correction must be done. However, for the upper stages, the images are sampled for micrometer particles. Thus, the smaller particles cannot be visible and the whole distribution for the stages cannot be obtained. As for the intermediate stages, this correction is possible as it will be seen in the nest paragraph. 3.4. Application to fly ash particles Once the method validated, it was applied to fly ashes particles from wood combustion. The size distribution of fly ash particles is concentrated on the smaller and intermediate stages. Thus, this analysis is restricted to stage 2 and stage 5 which respectively represent the study of a lower stage and an intermediate stage. Particles from stages 2 were studied (typical image Fig. 19 and distribution Fig. 20) and the average of the equivalent diam-

Fig. 19. SEM images of fly ash particles impacted on stage 2.

Fig. 20. Number size distribution of Fly Ash particles impacted on stage 2 (114 measures).

eters evaluated at 75 nm (Gaussian distribution fitting with a correlation factor of 98% and 10 nm bin size). Thus, the equivalent density for deq = 75 nm and dae = 76 nm is 1.0 g cm−3 . The variance for the equivalent diameter is 10 nm, then 0.9 ≤ ρ ≤ 1.2 g cm−3 . Fine particle losses on stage 2 can be considered as negligible (Fig. 18). Images sampled on stage 5 have been taken and the results of the measures show a very wide distribution (Fig. 21). According to the resolution of the SEM and the size distribution, it can be considered that the measures are representative of all the particles collected on the stage 5 (impacted and fine particles). The distribution obtained is very much spread out and the mean value found is deq = 80 nm if all the particles are considered. As a consequence, the density found with deq = 80 nm and dae = 324 nm is 7.3 g cm−3 . This value is not reasonable for such aerosols. However, according to the analysis of the fine particles losses for this experience (Fig. 18), with 70% of fine particle losses, the distribution of Fig. 21 can be truncated to keep the bigger 30% which leads to a 216 nm equivalent diameter and the density is re-evaluated to 1.9 g cm−3 . In conclusion, the density found for the particles impacted on stages 2 and 5 is rational for such an aerosol [13]. Then, according to the results, this method meets the expectations: firstly, the values of ρ are in the expected range, secondly the

Fig. 21. Number size distribution of Fly Ash particles impacted on stage 5 (211 measures).

N. Coudray et al. / Precision Engineering 32 (2008) 88–99

method gives ρ for each stage with an acceptable precision (less than 15%). 4. Conclusion In this paper, a method is proposed to evaluate the density of an aerosol sampled with the ELPI. This value is essential for the distributions (size, number, mass. . .) evaluated by the impactor, but can often not be measured with other usual methods. The first benefit of this study is that a new substrate was tested: glass substrates permit the impaction of the particles with the ELPI and is suited to scanning electron and Optical microscope observations. Also, according to the SEM observations of the Silica particles before and after the sampling, the substrate does not seem to alter the aspect of the particles. The other benefit of optical microscope imaging is that it permits to select the regions of interest and thus facilitates and speeds the SEM study. General views of the stages improve the comprehension of the ELPI and the impaction phenomena, putting in view a new kind of losses (turbulences). Then, the treatment of the SEM images is quite straightforward with classical image processing tools. Thus, several images can easily be treated to extract an equivalent diameter distribution. Furthermore, other geometrical parameters can be extracted (Feret’s diameter, bounding rectangle. . .) if the complexity of the particle shape requires it. The average equivalent diameter was extracted by fitting a Gaussian to the distribution for the studied stage, and the correlation factor validates the approximation. Based on the mean deq and dae , ρ was evaluated and the method validated with Silica particles, the density of which was previously measured with a pycnometer. The results are specially exploitable for the lower and upper stages. Applied to fly ashes, results are promising, ρ standing between 0.87 and 2, depending on the stage. However, when the size of the particles is widespread, all the particles cannot be acquired on a same picture because of the resolution of the microscope. This problem occurs mostly for the upper stages, causing a truncated distribution. When a wide range can be observed and measured, the result is still biased if the differentiation between the lost and impacted particles is not done. Evaluation of the different losses leads to a better approximation of ρ. Indeed, rebounds and fine loss particles are well known as being limitations of the ELPI. The distribution may then not only be wide, but also bimodal (especially for the intermediate

99

stages), which wrongs the Gaussian model. In other cases, a lognormal distribution was observed. Then, another model to fit the distribution must be studied to enhance the precision of the method. Thus, this method permits an evaluation of ρ, even though the variation for its value is quite large. This variation would probably be reduced by improving the losses calculation and when a better model is applied for fitting the distributions. References [1] Shi JP, Khan AA, Harrison RM. Measurements of ultrafine particle concentration and size distribution in the urban atmosphere. Sci Total Environ 1999;235:51–64. [2] Tsukamoto Y, Goto Y, Odaka M. Continuous measurement of diesel particulate emissions by an Electrical Low-Pressure Impactor. SAE Technical paper series 2000-01-1138. 2000. [3] Lavta-Somppi J, Moisio M, Kauppinen E, Valmari T, Ahonen P, Keskinen J. Aerosol formation in Fluidized bed Incineration with waste sludge. J Aerosol Sci 1998;29:461–80. [4] Marjam¨aki M, Keskinen J, Chen DR, Pui DYH. Performance evaluation of the electrical low-pressure impactor (ELPI). J Aerosol Sci 2000;31:249–61. [5] Moisio M. Real time size distribution measurement of combustion aerosols. PHD Thesis, Tempere University of Technology publications 279, Tempere, Finland, 2000. [6] Boudier T. Elaboration d’un mod`ele de d´eformation pour la d´etection de contours aux formes complexes. Innov Techn Biol Med 1997;18(1):1–13. [7] Chang M, Kin S, Sioutas C. Experimental studies on particle impaction and bounce: effects of substrate design and material. Atmos Environ 1999;33:2313–22. [8] Marjam¨aki M, Keskinen J, Chen DR, Pui DYH. Effect of impaction plate roughness and porosity on collection efficiency. J Aerosol Sci 2003;35:301–8. [9] Huang CH, Tsai CJ. Influence of impaction plate diameter and particle density on the collection efficiency of round-nozzle inertial impactors. Aerosol Sci Technol 2002;36:714–20. [10] Yamamoto N, Fujii M, Endo O, Kumagai K, Yanagisawa Y. Broad range observation of particle deposition on greased and non-greased impaction surfaces using line-sensing optical microscope. J Aerosol Sci 2002;33:1667–79. [11] Chaucherie X, Mangenot C, Vatry G. D´etermination de la granulom´etrie des a´erosols dans les e´ missions diffuses d”ateliers sid´erurgiques: PM10, PM2,5, PM1 et PM0,1., ADEME-Rapport final RE109/a, dossier N◦ 802 0275, 2004. [12] Virtanen A, Marjam¨aki M, Ristim¨aki J, Keskinen J. Fine particle losses in electrical low-pressure impactor. J Aerosol Sci 2001;32(3):389–401. [13] Gurupira, T., Jones, C.L., Howard, A., Lockert, C., Wandell, T., Stencel, J.M. New products from Coal Combustion Ash: selective extraction of particles with density <2, 2001 International Ash Utilization Symposium, Center for Applied Energy research, University of Kentucky, Paper 44, 2001.