«Abstract Fully automated online measurement of the size distribution of limestone particles on conveyor belt is presented based on 3D range data ...»
Automated Online Measurement of Limestone Particle Size
Distributions using 3D Range Data
Matthew J. Thurley
Luleå University of Technology, Luleå, SE-97187 Sweden
Fully automated online measurement of the size distribution of limestone particles on conveyor
belt is presented based on 3D range data collected every minute during 13 hours of production.
The research establishes the necessary measurement technology to facilitate automatic control of
rock crushing or particle agglomeration processes to improve both energy eﬃciency and product quality. 3D data from laser triangulation is used to provide high resolution data of the surface of the stream of rocks. The 3D data is unaﬀected by color variation in the material and is not susceptible to scale or perspective distortion common in 2D imaging. Techniques are presented covering; sizing of particles, determination of non-overlapped and overlapped particles, and mapping of sizing results to distributions comparable to sieving. Detailed variations in the product sieve-size are shown with an abrupt change when the size range of the limestone particles was changed.
Keywords: image segmentation, particle size measurement, range data, classiﬁcation, particle delineation
1. Introduction In the mining and aggregate industries a great deal of eﬀort goes into measuring or estimating the size distribution of particulate material. One reason is that suppliers of particulate material are typically paid to supply a speciﬁc size range of material. For both industries there is also a key desire for energy eﬃciency and size quality that is relevant throughout the mining process in blasting, crushing, and aggregating processes.
Processing plants want to control their value-added products such as baked limestone and iron ore pellets. Fast feedback of the product size range would allow for control optimisations to achieve both energy eﬃciency and an optimized product size range.
Mine and quarry operators want to measure the particle sizing results of all of these activities but sieving is typically impractical as a routine assessment tool due to slow feedback, inconsistent measurement, and time consuming interruption.
As a result there is an opportunity for online, non-contact, fully automated machine vision systems for measurement of particle size that can provide the necessary accuracy and fast feedback to facilitate process control and allow automatic control optimisations for both product size Email address: firstname.lastname@example.org (Matthew J. Thurley) URL: www.ltu.se/staff/m/mjt?l=en (Matthew J. Thurley) Preprint submitted to Process Control November 15, 2010 and energy eﬃciency. This is relevant to a vast range of processes that modify particle size such as rock blasting, crushing or agglomeration or particles, and baking processes in ovens and kilns.
There are however, a number of sources of error relevant to techniques that measure only what is visible on the surface of a pile as follows;
Segregation and grouping error, more generally known as the brazil nut eﬀect , describes the tendency of the pile to separate into groups of similarly sized particles. It is caused by vibration or motion (for example as rocks are transported by truck or conveyor) with large particles being moved to the surface. It is advisable to measure at a point early on the conveyor before the material has been subjected to excessive vibration and segregation.
Overlapped particle error, describes the fact that many particles are overlapped (see ﬁgure 1) and only partially visible and a large bias to the smaller size classes results if they are treated as small non-overlapped and sized using only their visible proﬁle. This error can be overcome in piles of particulate material using classiﬁcation algorithms based on 3D range data  successfully providing 82% classiﬁcation accuracy .
Figure 1: Illustration of overlapped and non-overlapped particles.
Capturing error, describes the varying probability based on size, that a particle will appear on the surface of the pile. In simple terms, the larger a particle is, the more likely one is to be able to see some part of it on the surface. For example, if a single particle is as large as the height of the pile of material, then it will always be visible, whereas a very ﬁne particle is almost certainly not visible. Thurley  has explored capturing error in laboratory rock piles but it remains a source of error in this application.
Proﬁle error, describes the fact that only one side (a proﬁle) of an entirely visible particle can be seen making if diﬃcult to estimate the particles size. However, if the particle is not overlapped, best-ﬁt-rectangle  has been demonstrated as a suitable feature for size classiﬁcation based on the visible proﬁle, that correlates to sieve-size.
In addition to these errors, we note that size measurement using imaging identiﬁes how many particles are observed of various size classes, but manual sieving measures the weight of particles in each size class. Therefore it is necessary to have a method of mapping from numbers of particles to weight of particles in order to provide a measurement of size that industry understands and can use. A technique based on physical sampling and sieving results is used in this work and explained in detail subsequently.
Particle size measurement using vision has been the subject of research and development for over 25 years  with a legacy of predominantly photographic based systems with widely varying degrees of success and no general solution available on the market.
Photographic based 2D imaging systems are subject to bias due to uneven lighting conditions, excessive shadowing, color and texture variation in the material, and lack of distinction between overlapped and non-overlapped particles.
In their review of a commercial photographic based 2D system Potts and Ouchterlony [7, pg. vi, viii] report that for their application the system erroneously assumes the resultant size distribution is unimodal and they conclude by expressing strong reservations saying 2D “imaging has a certain but limited usefulness when measuring the fragment size distribution in a muckpile or from a belt in an accurate way. It could probably detect rough tendencies in fragmentation variations, if the lighting conditions do not vary too much, and if cover glasses for camera lenses are kept clean”.
There are a number of publications relating to 3D size measurement, Noy [8, rocks], Frydendal and Jones [9, sugar beets], Kim et al. [10, river rock] Lee et al. . However, Frydendal Frydendal and Jones , and the presenting author  are the only publications (2D or 3D) to remove the bias resulting from overlapped particles. For conveyor belt applications some publications recommend installing a mechanical vibration feeder [10, 11] to separate rocks and prevent particle overlap. If there exists the possibility to install a vibration feeder at full-scale in the process it would simplify image analysis, and may improve results but with the added cost and maintenance of the additional equipment.
We note also that it is possible to automatically sample particulate material and separate the material into a single layer for imaging, however this requires additional material handling and additional equipment such as conveyors and vibrating feeders. The focus of this research is on systems for on-line measurement of the particle ﬂow that are non-contact, that is they do not require any additional material handling. If non-contact measurement is required then one must consider overlapped particle error and account for overlapped and non-overlapped particles.
Furthermore, in some circumstances, such as examination of rocks in in-production excavators , there is no other option than to account for overlapped and non-overlapped particles.
Frydendal  used graph theory and average region height to determine the entirely visible sugar beets but this relied on the regular shape and size of the beets. Only the presenting author has made this distinction between overlapped and non-overlapped particles using the advantages of 3D range data and in a manner that does not presume constraints on size or shape .
We use an industrial measurement system on conveyor belt based on laser triangulation (a projected laser line and camera at an oﬀset angle) collecting highly accurate 3D proﬁles of the laser line at about 3000 Hz. This high speed ensures we have a high density of 3D point data at a spacing between consecutive points in the direction of the belt of approximately 1 mm as the belt is running at 3 m/s. The imaging system is installed at a limestone quarry on the conveyor belt used for ship loading and measures the material on the belt during loading every minute. Figure 2 shows some images of the installed laser and camera system.
Using the same laser triangulation technology it is also possible to collect measurement points at 0.1 mm spatial resolution and 0.005 mm depth resolution as we have done on steel slabs (not yet published). Using such a setup it should be possible to conﬁdently measure particles with a lower size limit of 1 mm and be able to detect overlapped and non-overlapped particles on Figure 2: Installation of the measurement hardware above the conveyor belt.
conveyor belt. An equivalent reduction in measurement speed (one tenth conveyor belt speed) is necessary such as 0.3 m/s but it is theoretically possible to improve this using higher power lasers (beyond class 3B) but we have not tested this.
The presented analysis algorithms in this paper are not dependant on the laser triangulation measurement technology. Any other technique for capturing 3D surface data of a particle pile, such as stereo photogrammetry or time-of-ﬂight 3D cameras could also be used.
The computational speed of the analysis process is approximately 53 seconds on a 2 GHz Mobile Pentium 4 processing a data set of 590,000+ 3D points (2 m long section of the belt).
Furthermore, multi-core CPUs oﬀer an almost linear increase in the rate at which data could be sampled from the conveyor and processed. Further advances in both algorithmic eﬃciency and hardware are both available to improve computational time as the presented morphological image processing operations are readily parallelisable. Computation time could be reduced down to the order of few seconds or less for rapid automatic control applications. Once the analysis can be performed in a under a second then it will to become possible to measure and analyse the entire surface of the material passing under the measurement system.
There are a number of factors to consider regarding frequency of measurement and the length of the measurement sample. Computational speed is linearly dependant on the number of measured 3D points and therefore deﬁnes how fast measurement samples can be analysed, but it is also necessary to consider what is the largest particle size that should be measured and how many of these particles need to be measured to get a statistically valid measurement size. Given the circumstances in this application the 2 m long measurement sample is more than suﬃcient given the material top size of approximately 90 to 100 mm. Furthermore, the measurement time of under a minute is again suﬃcient given that ship loading takes place over several hours and it is suﬃcient to detect and conﬁrm particles out of the expected size range within several minutes.
2. Research Background
The presented research builds upon a series of achievements and research developed on both laboratory rock piles and industrial application.
We have previously implemented an industrial measurement system on conveyor belt for iron ore pellets  using the same laser triangulation measurement technology. The high speed camera system ensures we have a high density of 3D point data at a spacing between consecutive points in the direction of the belt of approximately 0.5 mm. This high data density has at least two advantages. Firstly it allows us to detect small sliver regions or crescent-like regions of overlapped particles and ensure that they are not merged into other regions. And secondly, it has ensured that we could detect a very high resolution when it came to measuring the size of each iron ore pellet allowing a size distribution with very ﬁne spacing of 5, 8, 9, 10, 11, 12.5, 13, 14, and 16+ mm size classes.
One of the key criteria for particle size measurement is therefore high data density as it deﬁnes the capacity to detect small overlapped particles, the lower limit on particle size that can be reliably detected, and the resolution of size classes detectable.
Iron ore pellet production is a process with cyclic feedback and surging. Changes in the rotational speed of the pelletising disk tend to be observed on the conveyor belt after 5–10 minutes, so measurement every minute for the pellet system  is more than adequate.
In addition we have performed a demonstration project for size measurement of rocks in underground LHD excavator buckets . A 3D vision system based on laser scanners was installed on the tunnel roof in a production area of an underground iron ore mine with 3D surface data of the bucket contents being collected as the LHD unit passes beneath. The project successfully demonstrated fragmentation measurement of the rocks in the bucket, identifying overlapped rocks, non-overlapped rocks, areas of ﬁne material, estimating the sieve-size of the visible rocks only, and the calculating the proportion of the surface that was identiﬁed as ﬁne particles below the observable resolution of the laser scanner.
3. Analysis of 3D Range Data
In the presented application for limestone particles the lower limit of particle size is 10 mm and 3D data points are collected at a spatial resolution of 1 mm.
The following analysis strategies apply morphological image processing techniques. These techniques are particularly useful for identifying spatial structure in image data. Morphological image processing operations typically apply a structuring element, which is a spatial probe similar to a neighbourhood operation, to the image. The reader is referred to the text Hands-On Morphological Image Processing  for a very clear explanation of this topic, or to any other good text on image processing if they require more information on these morphological image processing operations.