Image-based Method for Measuring Pellet Size Distribution in the Stable Area of Disc Pelletizer

Abstract

Disc pelletizer is widely used in the agglomeration process to form powdered iron ore into iron ore green pellets. The pellet size distribution (PSD) is one of the major measures of product quality. An imaging system is developed for measuring PSD in the stable area of the disc pelletizer. Image pre-processing is performed to extract the densely distributed pellets as foreground, followed by identifying surface pellets using pellet markers and K-means clustering method. Then a marker-controlled watershed algorithm is applied to segment overlapping pellets in the image. The pellet sizes and thus PSD are then obtained using circumscribed circle fitting. The proposed image-based measuring method was tested in a steel company. The measured PSD was compared with manual sieving results. It shows good accuracy for different size ranges under different pelletizing conditions. The proposed method also makes it possible to evaluate online the pellet quality.

1. Introduction

Disc pelletizer is widely used in steel manufacturing industry in the agglomeration process to form powdered iron ore into iron ore green pellets. The raw material is fed into the disc and sprayed by water (Fig. 1). With continuous rotation of the disc, fine particles are gathered and formed into larger pellets in the stable area, finally falling out of the disc as green pellets. The green pellets are then heat hardened to get the required properties in induration process and finally be fed into furnace. The pellet size in the disc should be constantly monitored to guarantee that it is in a desired size range, otherwise, the thermal process in the furnace will be greatly affected, leading to low heat transfer efficiency or bad product quality.^1,2,3)

Fig. 1.

Picture of a disc pelletizer.

Optical imaging is an efficient tool for measurement of particle size because it has advantages of simple hardware configuration and abundant choice of image processing algorithms. It has been used for size measurement of various types of particles, such as granules,^4,5) rocks,^6,7) bubbles^8,9,10) and coals.¹¹⁾ In processing of particle images, the main difficulty lies in segmentation of overlapping particles. The watershed transform to the gradient or edge image is popularly applied in literatures^6,7,8,9) to achieve the segmentation of overlapping particles, but it easily leads to severe over-segmentation due to large numbers of local minima/maxima on gradient image or distance transform map.¹²⁾ Over-segmentation results in many particles being measured smaller than they are. Chen et al.¹³⁾ proposed two image-based methods for size distribution measurement of nickel pellets using a video camera. In order to remove the overlapping effect, the Gaussian process regression (GPR) models and a counting rule were applied for gray-scale sub-curves and estimating the pellet diameter. The methods assumed that the pellets are free-falling with identical initial speed, which is however not the case for green pellets in disc pelletizers. For size measurement of green pellets, Harayama et al.¹⁴⁾ used a camera to capture images of the pellets inside the disc. To avoid the difficult problem of segmenting overlapping pellets in the image, this method only estimated the average diameter of the pellets by calculating the power spectrum of images, however, the PSD of the pellets was not measured. Thurley et al.^15,16) developed an imaging system to capture 3D surface data of pellets on the conveyer belt, with which PSD can be measured. However, as pointed out by Heydari et al.,¹⁷⁾ a conveyer belt often transports green pellets from various disc pelletizers, it is thus impossible to distinguish the green pellets from different discs, and the measured pellet size cannot lead to recognition of the malfunctioning disc. Heydari et al.¹⁷⁾ used a 2D camera to capture images of green pellets that falling directly from the outlet of the disc. The images were then processed using morphological methods, watershed transform and linear searching. A support vector machine (SVM) was then employed for classification of segmented pellets. This method is effective for images with sparse pellets. Images with dense pellets and thus strong overlapping effect, however, are ignored in the method. To what extent the measured PSD of sparse pellets can represent the PSD of all pellets, needs to be further investigated.

In the present work, we propose an image-based method for measuring PSD in the stable area of an industrial disc pelletizer. Its features are:

(i) Compared to imaging pellets on the conveyer, the proposed method captures images of the pellets inside the disc (the stable area) and thus can provide process information with less lag for control system of the pelletizing process.

(ii) Compared to pellets at the disc outlet, pellets in the stable area are more densely distributed, thus the measured PSD is more representative.

(iii) The image captured from stable area is full of overlapping pellets. In order to solve the difficult problem of segmenting overlapping pellets, pellet markers¹⁸⁾ and K-means clustering¹⁹⁾ are applied in the present work to identify the pellets exposed to surface, based on which a marker-controlled watershed algorithm^18,20) is then developed to segment overlapping pellets and measure the pellet size distribution with good accuracy.

The rest of this paper is organized as follows: in section 2, the disc pelletizer in a steel company and the imaging system are introduced; in section 3, the framework of the proposed method and the main image processing algorithms are described; in section 4, the measured PSD results are compared with manual sieving results, and the results obtained at different operating condition of the disc are also analyzed; In section 5, conclusions are made and future work are discussed.

2. Equipment and Imaging System

The equipment and imaging system is illustrated schematically in Fig. 2(a). The disc pelletizer is located in a steel company. It has a diameter of 6 meters and an inclination angle of 45 degree, operated with a rotation speed of 8 rpm. An industrial camera (Baumer VCXG-53M with a lens 50 mm, resolution 2048×2592) is used to capture images of the stable area in the disc. To reduce image distortions, the camera is so positioned that it is approximately perpendicular to stable area. Two LED lamps (250 watts each) are used to improve illumination. Each lamp consists of five LED lights arranged in an X-shape. To reduce the shadows in the image, the lamps are adjusted so that the light direction is nearly perpendicular to the stable area. Figure 2(b) is a sample image of the stable area. Result of camera calibration shows that the length of 1 mm in reality corresponds to 4.4 pixels in the image. It can be observed that the illumination in the middle of the image (region BFGC in the Fig. 2(b)) is much stronger than in other regions (AEFB and DHGC). To facilitate image segmentation, the region BFGC is processed separately, while image regions AEFB and DHGC are processed together.

Fig. 2.

(a) Illustration of the imaging system; (b) Raw image of the stable area.

3. Method

The proposed method for measuring the pellet size is given in Fig. 3. It consists of five main steps: (i) reducing noise in the image using Gaussian filter; (ii) extraction of the foreground by threshold segmentation; (iii) identification of surface pellets using pellet markers and K-means clustering; (iv) segmentation of overlapping pellets by marker-controlled watershed; (v) measuring the surface pellet size by circumscribed circle fitting. In the following texts, each step will be described in detail.

Fig. 3.

Flow diagram of the proposed method.

3.1. Gaussian Filtering

To reduce noise in the captured raw image I (x, y) that is generated during the camera imaging, a two-dimension Gaussian filter is applied in the preprocessing stage and obtain the filtered image I_g (x, y). The result is shown in Figs. 4(a)–4(b) for example.

Fig. 4.

An example illustrating main steps of the proposed method: (a) An example region R1 of the raw image Fig. 2(b); (b) Image after Gaussian filtering; (c) Extraction of the foreground; (d) Pellet markers detected by regional maximum method; (e) Identified surface pellets by K-means clustering; (f) The markers labeling the exposed fractions of the partially buried pellets; (g) Combined markers of (d) and (f); (h) Segmentation result of overlapping surface pellets using marker-controlled watershed algorithm; (i) Fitting circumscribed circles to the surface pellets.

3.2. Extracting the Foreground

The grayscale histogram of the image region BFGC is illustrated in Fig. 5(a), fitted by smoothing spline method.²¹⁾ It has an obvious unimodal shape with a peak at the critical intensity θ. Figure 6 shows the 3D intensity profile of the pellets of Fig. 4(b), in which the height of the “mountains” is proportional to intensity values. Each pellet has pixels with the maximum gray intensity (“mountaintop”) near its center due to strong light reflection. The intensity then gradually decreases from the center to the pellet edge. Based on these features, it is reasonable to use the critical intensity θ in the histogram as threshold to extract the pellets as foreground:   

$$I_g^{*}(x,y)=\{\begin{array}{l}{I}_g(x,y),\text{ if }{I}_g(x,y)\ge \theta \\ 0,\text{ if }{I}_g(x,y)<\theta \end{array}$$(1)

Fig. 5.

(a) Grayscale histogram of image region BFGC; (b) Grayscale histogram of image regions AEFB and DHGC.

Fig. 6.

3D intensity profile of the pellets of Fig. 4(b).

The extracted foreground image $I_g^{*}(x,y)$ is shown in Fig. 4(c) for example. Similar procedure is applied to the image regions AEFB and DHGC. Due to the relatively weak illumination, the histogram of image regions AEFB and DHGC shown in Fig. 5(b) has its majority closer to the vertical axis and has a smaller intensity range than the histogram of region BFGC. The critical intensity θ′ used to extract the foreground of image regions AEFB and DHGC is smaller than θ after the histogram is fitted by the smoothing spline method.

3.3. Identification of the Pellets on Surface

In the image of the stable area, some pellets are fully exposed to surface, while some pellets are almost buried underneath and are thus unsuitable to be used for pellet size measurement. It is therefore necessary to identify those pellets exposed to the surface. For this aim, pellet marker and K-means clustering are used in the present work.

3.3.1. Pellet Markers

The grayscale regional maximum²²⁾ is a connected component of pixels all has a higher gray value than the pixels in the neighborhood. The “mountaintops” in Fig. 6 inspire us to detect the grayscale regional maximums as markers of the pellets. However, in some cases, the regional maximum involves only several pixels and could be caused by noises. Hence, we apply open operation of grayscale reconstruction with a disc-shape structuring element of radius 3 to eliminate those noises^20,23) and maintain normal markers. The detected pellet markers (bright patches) are shown in Fig. 4(d) for example.

3.3.2. Identification of Surface Pellets Using K-means Clustering

Due to direct exposure in the light, pellets on the surface generally have higher gray intensity than pellets buried underneath. It is possible to classify the surface pellets and the buried pellets using threshold method. However, a fixed threshold is difficult to obtain because no prior knowledge is available. Moreover, the variation of illumination in each image requires that the threshold should have good robustness. K-means clustering belongs to an unsupervised machine learning method and is advantageous in our case to find a robust threshold.

Denote the mean intensities of the detected pellet markers as {t₁, t₂,…, t_i}, we apply K-means clustering algorithm to classify the markers into two clusters: C₁ (markers of the surface pellets) and C₂ (markers of the buried pellets) by minimizing the within-cluster sum of square E.   

$$E=\sum _{j=1}^2{\sum _{t_i\in {\text{ C}}_j}\Vert t_i-u_j\Vert }^2$$(2)

where u_j is the mean gray value of cluster C_j. The markers of surface pellets are shown in Fig. 4(e) for example, with which the pellets on the surface can be identified.

3.4. Segmentation of Overlapping Pellets Using Marker-controlled Watershed Method

Observation of Fig. 4(c) shows that the roughly extracted foreground contains many overlapping pellets, i.e. overlapping between surface pellets, or overlapping between surface pellets and partially buried pellets. It is thus necessary to segment these overlapping pellets to avoid oversize in measurement. The accuracy of pellet labeling affects the segmentation of overlapping pellets. However, as shown in Fig. 4(d), the exposed fraction of some partially buried pellets are not labeled by markers because their regional maximum involves only several pixels and is treated as noises. This could result in under-segmentation, that is, many segmented pellets will be oversized, because the exposed fractions of buried pellets will be mistakenly identified as parts of their neighboring surface pellets in segmenting process. To solve this problem, the image processing steps described in Section 3.3 are repeated with a disc-shape structuring element of radius 1 and the markers in C₂ in Eq. (2) to label the exposed fractions of the buried pellets. The results are shown in Fig. 4(f) for example.

As already discussed in Section 3.2, the gray intensity of the pellet decreases monotonically from the interior to the boundary, which forms the grayscale topography and allows us to apply the marker-controlled watershed algorithm to directly segment the overlapping pellets. The segmentation is then proceeded as follows: the combined markers (see Fig. 4(g)) for the pellets and fractions of the partially buried pellets, are imposed as the regional minima in the complement of extracted foreground image $I_g^{*}(x,y)$; the watershed lines are then computed to separate the regional minima and segmentation is finally accomplished. The segmentation result is shown in Fig. 4(h) for example.

3.5. Measurement of the Pellet Size and PSD

With segmented pellets, the pellet size can be now measured. In literature, two methods have been used to approximate the shape of pellets: rectangle fitting¹⁵⁾ and circle fitting.¹⁷⁾ Since the pellets in the disc are nearly spherical and are sieved by round-hole mesh in practice, the circumscribed circle fitting (or smallest enclosed circle fitting²⁴⁾) is applied in the present work to approximate the pellet shape: (1) all pixels of the segmented surface pellet are scanned to find the leftmost pixel p_l, rightmost pixel p_r, uppermost pixel p_u, and bottommost pixel p_b; (2) circle fitting is performed based on the four pixels and a circle (denoted as mec({p_l, p_r, p_u, p_b})) is determined; (3) if all pixels of this pellet are inside or on this circle, this circle is then taken as the circumscribed circle of the segmented surface pellet; if there still exist pixels outside this circle, the pixel with longest distance to the circle center point (denoted as $p_f^1$) will be added in circle fitting to generate a new circle mec({p_l, p_r, p_u, p_b, $p_f^1$}); (4) step (3) is repeated n times until there are no pixels outside the fitted circle mec({p_l, p_r, p_u, p_b, $p_f^1$…$p_f^n$}) and this circle is finally taken as the circumscribed circle of the segmented surface pellet. The fitted circumscribed circles of surface pellets are shown in Fig. 4(i) for example.

In order to visually and comprehensively demonstrate the processing effect of the proposed method, the extracted more than 2000 surface pellets of the raw image Fig. 2(b) are shown in Fig. 7. The size of each pellet is then quantified using the diameter d of the fitted circle. The PSD can be obtained by statistical analysis of size of all surface pellets.

Fig. 7.

The extracted surface pellets of Fig. 2(b).

4. Results and Discussions

4.1. Experimental Validation

4.1.1. Experimental Procedures

The experiments were performed in a steel company. Images of the stable area in the disc pelletizer were automatically captured by the camera in the time period 11:32–14:32, during which the disc was running continuously and some process parameters were manually adjusted by the technical worker, as illustrated in Fig. 8. For example, there material feed was stopped from 12:57 until 13:02, and the water feed was adjusted at 13:03 and 13:06. The image acquisition interval is about 20 seconds, corresponding to about 530 image frames in total. (Note: the images from 12:57 to 13:00 are not available because the storage disc was full at that time).

Fig. 8.

Actions during image acquisition in the time period 11:32–14:32.

In order to obtain the actual PSD of the pellets, a shovel was used to manually collect some sample pellets from the disc. To avoid invasive collecting, the shovel was not inserted deeply into the stable area; rather, it was put against the surface of the stable area to collect pellets falling down from the stable area. The pellet collecting process was performed manually at different time at 11:33, 12:26, 13:00, 13:05, 13:07, 13:15, 13:18 (see Fig. 8). Each process took about 3 seconds.

Those collected sample pellets were then sieved manually using nine round-hole meshes with hole diameter of 6 mm, 8 mm, 9 mm, 10 mm, 11 mm, 12 mm, 13 mm, 14 mm and 16 mm (Fig. 9). The number of pellets (and thus percentage of the pellets) left on each mesh was counted, with which nine ranges of pellet size (6–8 mm, 8–9 mm, 9–10 mm, 10–11 mm, 11–12 mm, 12–13 mm, 13–14 mm, 14–16 mm, ≥16 mm) were obtained, as shown in Table 1. The number of small fragments of pellets and iron ore materials (<6 mm) were impossible to count and were therefore not used in analysis. The sieving results will be used to test the accuracy of the proposed image-based measuring method.

Fig. 9.

The round-hole meshes.

Table 1. The sieving results using round-hole meshes.

Sampling time	Percentage of pellets in the size range (%)
	6–8 mm	8–9 mm	9–10 mm	10–11 mm	11–12 mm	12–13 mm	13–14 mm	14–16 mm	>16 mm
11:33	9.7	10.4	14.3	16.7	15.1	14.0	9.8	7.1	2.8
12:26	10.2	11.2	16.0	17.4	13.1	13.4	8.8	6.8	3.2
13:00	3.7	1.4	3.2	1.4	3.2	4.3	10.6	31.9	40.2
13:05	26.2	16.9	21.6	11.9	8.5	6.6	3.7	3.0	1.6
13:07	18.4	17.5	22.5	16.6	12.2	7.1	3.5	1.3	0.9
13:15	8.6	8.9	13.5	14.9	16.8	15.1	10.1	8.9	3.2
13:18	13.9	9.2	15.2	16.7	12.3	13.3	10.0	6.3	3.0

4.1.2. Comparison of Image-based Measuring Results and Sieving Results

The size distributions measured by the imaging method are then compared to the sieving results. The comparison result at 11:33 for example is shown in Fig. 10. Good agreement is found in eight of all the nine size ranges of pellets. The biggest deviation occurs in the size range 6 mm–8 mm. This could be caused by the manual pellet collecting method used in experiments: the shovel was put against the surface of the stable area and collected pellets falling down from the stable area; due to high velocity of the rotating disc, it was inevitable that some smaller pellets under the pellets surface were also thrown into the shovel; in contrast, the imaging system captured the upper layer pellets of the stable area, where very small pellets were mostly buried under bigger pellets due to segregation effect.²⁵⁾

Fig. 10.

Size distribution measured by the proposed imaging method and comparison to sieving results at time 11:33.

For the disc pelletizer in the present work, the desired pellet size is 9 mm–14 mm (good product quality). By use of the proposed method, the size distribution at different time can be measured automatically. In Fig. 11, the cumulative percentages of pellets measured by the proposed method are compared with sieving results for all the seven tests at different time. It can be seen that the agreements between the imaging results and sieving result are reasonably good even when the disc pelletizer was operated under different conditions. Both imaging results and sieving results show that: for the four cases (11:33, 12:26, 13:15, 13:18) where the material feed and water feed were generally stable (see Fig. 8), the pellet quality was satisfactory because only about 20% pellets were smaller than 9 mm and about 10% pellets were bigger than 14 mm; due to lack of material feed from time 12:57 to 13:02 (see Fig. 8), about 70% of the pellets sampled at time 13:00 exceeded the upper size range d ≥ 14 mm (bad product quality); the frequent manual adjustment of the water feed in the time range 13:03–13:06 also resulted in bad product quality at 13:05 and 13:07, where at least 30% of the pellets had a diameter less than 9 mm.

Fig. 11.

Comparison of imaging results and sieving results at different time.

4.2. Monitoring of Pellet Quality

As demonstrated in Section 4.1, the proposed imaging method is effective in measuring the size distribution of the pellets in the stable area. With the method, it is also possible to monitor the pellet quality, as shown in Fig. 12 where the percentage of pellets with good quality (size requirement: 9–14 mm) are monitored online in the time period 11:32–14:32. It can be seen from this figure that, 70%–80% of the pellets meets the size requirement when the disc is in normal operation. However, there are also cases in which the disc pelletizer produces pellets of bad quality, for example at time 13:00 where only 29% of the pellets is in the desired size range. Note that the pellet quality is monitored based on images captured from the stable area, instead of images from the disc outlet. Thus, based on the measured PSD as feedback signal, the automatic control system will have enough time to adjust the operating parameter (water or material feed) to avoid worse product quality.

Fig. 12.

Online monitoring of pellet quality using the proposed method.

5. Conclusions and Future Work

The present work provides an image-based method for measurement of the pellet size distribution in disc pelletizers. The effectiveness of the imaging system and the proposed image processing algorithms was tested on an industrial disc pelletizer in a steel company. It can be concluded that:

(1) The pellets exposed to surface can be identified using pellet markers and K-means clustering, based on which the overlapping pellets can be well segmented by marker-controlled watershed;

(2) The proposed method is effective in measuring pellet size distribution. The results are in good agreement with sieving results for different size ranges, even when the disc was operated under different conditions.

(3) Different to most work reported in literature, the proposed method uses images captured directly from the stable area within the disc (the pellets there are densely distributed), instead of images captured at the disc outlet (the pellets there are relatively sparse) or images on the conveyer. This brings the advantages that the measured PSD is more representative and has less time-lag as feedback signal to the control system.

The proposed method can be used for online monitoring of pellet quality in the disc. However, like other optical imaging methods, it also has limitations: the illumination affects the image quality. Before the proposed method is applied, an appropriate light source should be selected according to the specific working environment of the disc pelletizer.

As the disc pelletizer is the main equipment for powder pelletizing and can be used for pelletizing of various raw materials in the industries of compound fertilizer, coal, metallurgy and cement, it gives a very large industrial application prospect of the proposed method.

Acknowledgements

Financial support from National Natural Science Foundation of China (No. 61374149) and Hunan Key Laboratory of Intelligent Robot Technology in Electronic Manufacturing (No. 2018001) are greatly appreciated.

References

• 1)   T.  Higuchi,  N.  Takeuchi,  Y.  Ishigaki,  T.  Yamamoto,  H.  Matsuno and  N.  Oyama: Tetsu-to-Hagané, 103 (2017), 289 (in Japanese).
• 2)   S.  Dwarapudi,  T. U.  Devi,  S. M.  Rao and  M.  Ranjan: ISIJ Int., 48 (2008), 768.
• 3)   T.  Umadevi,  P.  Kumar,  N. F.  Lobo,  M.  Prabhu,  P. C.  Mahapatra and  M.  Ranjan: ISIJ Int., 51 (2011), 14.
• 4)   P.  Facco,  A. C.  Santomaso and  M.  Barolo: Chem. Eng. Sci., 164 (2017), 246.
• 5)   N.  Sandler: Int. J. Pharm., 417 (2011), 227.
• 6)   M. J.  Thurley: J. Process Contr., 21 (2011), 254.
• 7)   G.  Christie,  K.  Kochersberger and  A. L.  Abbott: Mach. Vis. Appl., 26 (2015), 31.
• 8)   Y. M.  Lau,  N. G.  Deen and  J. A. M.  Kuipers: Chem. Eng. Sci., 94 (2013), 20.
• 9)   A.  Karn,  C.  Ellis,  R.  Arndt and  J.  Hong: Chem. Eng. Sci., 122 (2015), 240.
• 10)   J.  Liu,  W.  Gui,  Z.  Tang,  C.  Yang,  J.  Zhu and  J.  Li: Miner. Eng., 45 (2013), 128.
• 11)   S.  Al-Thyabat,  N. J.  Miles and  T. S.  Koh: Miner. Eng., 20 (2007), 72.
• 12)   H.  Yang and  N.  Ahuja: Pattern Recogn., 47 (2014), 2266.
• 13)   J.  Chen,  J.  Yu and  Y.  Zhang: Comput. Chem. Eng., 64 (2014), 13.
• 14)   M.  Harayama and  M.  Uesugi: Proc. 1992 Int. Conf. on Industrial Electronics, Control, Instrumentation, and Automation, IEEE, Piscataway, NJ, (2002), 1613.
• 15)   M. J.  Thurley and  T.  Andersson: Miner. Eng., 21 (2008), 405.
• 16)   T.  Andersson and  M. J.  Thurley: Powder Technol., 206 (2011), 218.
• 17)   M.  Heydari,  R.  Amirfattahi,  B.  Nazari and  P.  Rahimi: Powder Technol., 303 (2016), 260.
• 18)   S.  Lefèvre: Int. Conf. on Computer Analysis of Images and Patterns, Springer-Verlag, Berlin, (2007), 579.
• 19)   C. M.  Bishop: Pattern Recognition and Machine Learning (Information Science and Statistics), Springer-Verlag, Berlin, (2006), 424.
• 20)   L.  Vincent: IEEE Trans. Image Process., 19 (1993), 176.
• 21)   Y.  Wang: Monographs on Statistics and Applied Probability, Vol. 121, CRC Press, Boca Raton, FL, (2011), 11.
• 22)   P.  Soille: Morphological Image Analysis: Principles and Applications, Springer-Verlag, Berlin, (2003), 166.
• 23)   P. W.  Huang and  Y. H.  Lai: Pattern Recogn., 43 (2010), 1550.
• 24)   E.  Welzl: New Results and New Trends in Computer Science, Vol. 555, ed. by H. A. Maurer, Springer-Verlag, Berlin, (1991), 359.
• 25)   A.  Rosato,  K. J.  Strandburg,  F.  Prinz and  R. H.  Swendsen: Phys. Rev. Lett., 58 (1987), 1038.