You are on page 1of 11

Validation of a new cross-belt cutter for sampling gold ores

A.C. Chieregati1, E.A. Amaral Jr.2 and J.C.O. Souza3

1 Professor, Department of Mining and Petroleum Engineering; University of São Paulo, 05508-030, São
Paulo, SP, Brazil;
2 Plant Manager, Jacobina Mineração e Comércio, Yamana Gold Inc., 44700-000, Jacobina, BA, Brazil;
3 Process Engineer, Jacobina Mineração e Comércio, Yamana Gold Inc., 44700-000, Jacobina, BA, Brazil;

Cross-belt samplers are designed to sample particulate material in processing plants usually when there is no
space to install a conventional linear falling stream sampler. Most of the available cross-belt samplers are
mounted on the top of conveyor belts and remove each increment from a loaded, moving conveyor with a 360-
degree rotation of an open-faced cutter in a plane perpendicular to the material flow. Despite the practicality
of this equipment and the fact that it occupies very little space in the plant, cross-belt samplers are well known
for collecting biased samples, especially due to the increment extraction error (IEE). Sampling consultants
around the world tend to condemn this sampling equipment, claiming that it is incapable of colleting the fine
material from the bottom part of the load, and, since the sampling probability is not uniform for all fragments
and depends on its position in the stream, this kind of sampler generates biased samples and its use should
be avoided. Furthermore, the greater the segregation effect between particles, the larger the bias. However,
some plants, whose initial design did not predict the installation of a falling stream sampler, must rely on cross-
belt samples for process control and reconciliation. For the past years, sampling equipment manufacturers
have been putting effort in modifying the cross-belt sampler design in order to comply with the principles of a
correct sampling. Changes to the cutter material, shape, position, speed and movement in relation to the
material flow have been made and are currently being tested for different types of ore with various chemical e
physical characteristics. This paper presents the continuous improvements made to the cross-belt sampler at
Jacobina, a gold mine of Yamana Gold in the northeast of Brazil, and describes recent tests performed for the
validation of the new cutter installed at the plant feed.
Keywords: Cross-belt samplers. Gold. Sampling.

Cross-belt samplers are known for their bad reputation in minerals processing plants. Lyman et al. (2010)
carried out many experiments for bias testing cross-belt samplers, such as the t-tests and the Hotelling T
squared tests, concluding that the cross-belt sampler is always biased compared to belt cuts, because it is
incapable of collecting the material representing the complete cut on the belt.
However, some suppliers have developed a modified cross-belt cutter designed to operate in a way that
minimises the disturbance to the non-sampled material, producing more correct cuts and more representative
samples. Multotec, with collaboration with Dr. Dominique François-Bongarçon, named it the True-Belt®
Sampler and ensures that the new design complies with international sampling standards, guaranteeing the
structural absence of sample biases. Robinson, Sinnott and Cleary (2010) conducted several DEM simulations
using the modified cutter to investigate the mechanisms that might lead to sample bias in order to help coal
industry personnel to make better decisions about the use of cross-belt samplers.
This paper presents the new validation study for the modified cross-belt cutter installed on the flow that feeds
a ball mill at Jacobina. After validating the new cutter with a 20 paired data set (Chieregati et al., 2017), new
tests were performed to confirm the results for richer gold ores. The sampler was designed by Engendrar and
is shown in Figure 1, following the recommendations presented by Robinson, Sinnott and Cleary (2010). The
cutter is positioned with an approximate angle of 33.7° in relation to the belt in a way that the resultant of the
cutter speed and belt speed vectors is a vector of approximately 73.4°, almost perpendicular to the belt, as
shown in Figure 2. To generate such a vector, the cutting speed should be 1.5 times the speed of the conveyor
belt and the cutter should move against the flow.

FIG 1 – Modified cutter for cross-belt samplers and resulting vector (courtesy of Engendrar).

FIG 2 – Cross-belt cutter movement on the belt (modified from Robinson, Sinnott and Cleary, 2010).


Sampling and sample preparation procedures

Jacobina is an underground gold mine owned by Yamana Gold and located in Bahia, in the northeast of Brazil.
The auriferous mineralization currently mined is found in the conglomerates of metasedimentary rocks of the
Serra do Corrego Formation. The ore bodies, called reefs, are characterised by strata of predominantly

conglomerates with higher concentrations of gold located at the top of the layer. Total proven and probable
mineral reserves are estimated as 2.0 million oz with an average gold grade of 2.86 g/t.
For more than five years the operation has continuously improved sampling and sample preparation
procedures, performing sampling tests and optimising sampling protocols and equipment. For determining the
sampling constants K and , both the heterogeneity test and the sampling tree method were carried out.
Results showed a slight difference for the constants K and a larger difference for . Table 1 presents the
fundamental sampling error (FSE) for the optimised sampling and sample preparation protocol based on the
results of the heterogeneity test, where K = 274.12 and  = 1.7278. Figure 3 illustrates the same protocol as
a flowchart. Although most of the gold particles are fine, the deposit also presents coarse gold, therefore, gold
grades are estimated by the weighting average of standard 50-g fire assay triplicates to minimise the nugget

TABLE 1 – Fundamental sampling error for each stage of the optimised sampling protocol at Jacobina’s lab.

initial final d IHL rel var rel dev

mass (g) mass (g) (cm) (g) (s2FSE) (sFSE rel)
1. Primary sampling 50,000,000 18,000 1.0 274.12 0.015223 12.34%
2. Crushing 18,000 18,000 0.10 5.13 0.000000 0.00%
3. Splitting 18,000 500 0.10 5.13 0.009976 9.99%
4. Pulverization 500 500 0.0104 0.10 0.000000 0.00%
5. Selection of analytical sample 500 50 0.0104 0.10 0.001849 4.30%
TOTAL (s FSE) 0.027048 16.45%

FIG 3 – Optimised sample preparation protocol at the laboratory.

Based on previous consultant recommendations and aiming to improve the reconciliation between the mine
and the mill, personnel decided to carry out a bias test on the mill feed, where a standard cross-belt sampler

was installed. Before starting the test, some measures were taken, such as replacing the sampler cutter with
the modified cutter of Figure 1, replacing the old cutting blades with new ones (Figure 4) and calculating the
minimum sample masses for both grade (Equation 1) and particle size distribution analysis (Equation 2).

K d 274  11.73
MS  2   10,703 g  10.7 kg (1)
s FSE 0.162

d3 13
M S  18 f   18  0.5  16   5,625 g  5.6 kg (2)
s 2FSE 0.162

Where MS is the minimum sample mass, f is the shape factor,  is the gold alloy density, d is the top size of
the particles, and sFSE is the maximum relative standard deviation of the fundamental sampling error.

FIG 4 – Old and new cutting blades.

The new cutting blades shape was, then, optimised by staining them with oil and verifying whether the blades
touched the belt or not after the cut.

Particle size distribution analysis

The first test carried out to check the capability of the cross-belt sampler to collect unbiased samples was the
comparison between particle size distributions of both stopped belt and cross-belt samples. In the previous
optimisation work (Chieregati et al., 2017) three paired samples were collected at the cross-belt discharge
point and on the stopped belt just after the cross-belt cut. For the present study, two more paired samples
were collected at the same points. The cross-belt samples were represented by one whole cross-belt cut, while
the stopped belt samples were represented by 1-m cuts of particulate material on the belt, carefully collected
using a frame. The sample masses varied from 9 kg to 24 kg approximately, respecting the minimum sample
mass for size distribution analysis. All samples were screened between 3/8” and 400# and the gold grades
were determined for each size fraction.

Bias test
The second test carried out at Jacobina’s plant was the bias test, where 35 paired samples were collected at
the cross-belt discharge point and on the stopped belt. The stopped belt samples were represented by 1-m
cuts on the belt, which were taken carefully and whose masses varied from 22 kg to 24 kg, characterising the

unbiased method. The cross-belt samples were represented by the whole cross-belt cut, whose masses varied
from 6 kg to 11 kg, characterising the method to be checked. The conveyor belt was stopped right after the
cross-belt cut and the samples were taken close to the respective cut.
There are statistical tests for bias verification between pairs of samples, whose theoretical foundations and
experimental procedures are described in international standards (Standards Australia, 2003; International
Organization for Standardization, 1991). In the experimental methods given in these international standards,
the results obtained from the method to be checked (referred to as “Method B”) are compared with the results
of a reference method (referred to as “Method A”), which is considered to produce practically unbiased results.
When the difference between the results obtained from Method B and those obtained from Method A is not
considered statistically significant, Method B may be adopted as the routine method.
The reference method for checking the bias is the stopped belt method. Although the standards do not specify
the procedures to collect the samples and how to calculate the minimum sample masses, the authors used
Pierre Gy’s Theory of Sampling (Gy, 1992; Pitard, 1993) to guarantee that the sample collected on the stopped
belt would be correct (unbiased), and to calculate minimum sample masses for both grade and size distribution
analysis, as previously presented.
Based on the standards, bias is assessed by application of the t-test (one-sided) at the 5% significance level,
by determining whether the difference between the results of Method A and of Method B are due to random
variations or are statistically different. The minimum number of paired sets of measurements is 20, however,
it depends on the standard deviation of the differences based on the data set and the absolute value of the
bias, , to be detected. This method may also be applied for checking a possible significant difference in the
results obtained from samples of one lot collected in different places, such as loading and discharging points.
Method B is considered unbiased when |t0| < t. The value of t is related to the number of sample pairs, k (see
Appendix 1), and t0 can be calculated as follows.

t0  (3)

Whered is the mean of the differences, di, between measurements obtained in accordance with Method A
and B, sd is the standard deviation of the differences and k is the number of paired data sets. The number of
pairs, k, is considered sufficient when the required number of pairs nr ≤ k. If nr ≥ k, additional experiments
must be carried out on nr-k data sets. The required number of data sets, nr, can be found in Appendix 1,
according to the value of the standardised difference, D.

D (4)

Where  is the magnitude of the bias to be detected and sd is the standard deviation of the differences.


Particle size distribution analysis

The graphs presented in Figure 5 show the particle size and the gold grade distributions for the cross-belt
samples compared to the stopped belt samples. The stopped belt sample masses varied from 23 kg to 24 kg,
and the cross-belt sample masses varied from 9 kg to 10 kg.

FIG 5 – Comparison of particle size and grade distribution between cross-belt and stopped belt samples.

Results show very similar particle size distributions and similar gold grade distributions. The more significant
discrepancies between gold grades are seen in the size factions with smaller retained masses, which was to
be expected, since smaller masses present higher standard deviation of the fundamental sampling error.
Furthermore, the 50-g fire assay technique can generate itself an error which is difficult to estimate.
However, the relative grade estimation error was small: 5.0% for the first paired sample (the gold grade
weighted average for the stopped belt sample was 3.302 g/t, and for the cross-belt sample was 3.467 g/t), and
4.2% for the second paired sample (the gold grade weighted average for the stopped belt sample was 2.561
g/t, and for the cross-belt sample was 2.668 g/t).

Bias test
The bias test described in the previous section was conducted on 35 paired samples taken on the stopped
belt (reference method or “Method A”) and by the cross-belt sampler (method to be checked or “Method B”).
Even though there is an average positive bias of 2.02% in favour of the cross-belt samples, the relative grade
estimation error using the 35 paired samples was only 1.72% and the statistical test shows no significant bias
between Method A and Method B, seeing that |t0| < t as per Table 2. The number of pairs was sufficient to
detect an absolute bias of 0.4 ppm (chosen based on a maximum relative standard deviation of the
fundamental sampling error of 16% and on the average gold grade of all samples), seeing that nr ≤ k.
Therefore, the modified cross-belt sampler may be adopted as a routine sampling method to estimate the
processing plant feed grade.

TABLE 2 – Bias test: Method A (stopped belt) × Method B (cross-belt sampler).


Pair Au (ppm) Absolute Relative
index difference di difference
1 2.59 1.94 0.648 33.33%
2 2.45 2.09 0.361 17.32%
3 1.73 2.32 -0.583 -25.18%
4 1.66 2.64 -0.979 -37.10%
5 3.19 2.05 1.146 55.97%
6 1.91 1.72 0.190 11.03%
7 2.40 1.79 0.609 33.95%
8 2.70 3.54 -0.837 -23.64%
9 3.18 2.30 0.880 38.27%
10 2.40 2.50 -0.096 -3.85%
11 2.56 2.76 -0.196 -7.10%
12 3.41 2.47 0.936 37.86%
13 3.32 3.67 -0.351 -9.56%
14 3.07 3.65 -0.581 -15.92%
15 1.83 3.44 -1.618 -46.98%
16 2.55 1.76 0.789 44.93%
17 1.75 1.53 0.211 13.75%
18 1.96 2.66 -0.700 -26.31%
19 3.31 2.72 0.595 21.91%
20 3.20 3.10 0.102 3.29%
21 3.39 3.59 -0.197 -5.49%
22 2.28 2.04 0.238 11.67%
23 2.69 3.50 -0.810 -23.13%
24 2.65 2.45 0.204 8.33%
25 3.58 3.73 -0.145 -3.89%
26 2.53 2.60 -0.063 -2.43%
27 2.31 2.05 0.264 12.88%
28 1.51 2.77 -1.265 -45.63%
29 2.59 2.33 0.258 11.07%
30 2.29 2.79 -0.493 -17.69%
31 2.83 3.10 -0.264 -8.53%
32 1.92 1.57 0.346 22.02%
33 2.60 2.14 0.457 21.36%
34 1.53 1.45 0.078 5.39%
35 1.45 2.11 -0.659 -31.29%
Mean 2.49 2.54 -0.044 2.02%
Variance 0.371 0.453 0.416 6.66%
k 35
Sum -1.526
SSd 14.133
sd 0.645
t0 see Equation 3 -0.400
t for 35 pairs (see Table 3) 1.691
UNBIASED? is Method B unbiased? YES |t 0 | < t
 (%) bias to be detected (sFSE = 16% relative) 0.40
D standardised difference 0.62
nr required number of pairs (see Table 4) 32
SUFFICIENT? was the number of pairs sufficient? YES nr ≤ k

The relative difference plot shown in Figure 6 indicates that the cross-belt samples tend to overestimate low
grades (<2.5 g/t) and to underestimate higher grades (>2.5 g/t). This result confirms the tendency observed in
the 20 sample pairs used for the first bias test (Chieregati et al., 2017) and should be taken into consideration
when feeding the plant with lower or higher grades.

FIG 6 – Relative difference plot between results: cross-belt  stopped belt (reference method).

However, this graph must be carefully analysed, and the authors recommend that the fire assay technique be
checked to verify the chemical analysis biases that may be generated.

As proved by Robinson’s DEM simulations, all cross-belt samplers will generate some bias, with material from
different positions across the width of the belt and from top to bottom of the load on the belt being unequally
represented in samples. The bias depends on the amount of segregation of the material on the belt, on the
grade distribution by size fraction, and on the cutter design and operational parameters as well. The question
is how to decide whether the likely amount of bias is acceptable or not.
This study described two tests that help estimating the bias of an optimized cross-belt sampler and checking
if it is statistically significant or not. The bias test approach is purely statistical; however, it allows detecting a
bias and can be compared with the size distribution analysis of paired data in order to confirm its results.
Knowing that the grades vary according to the size fraction, when the particle size distribution of a reference
method corresponds to the one of the method to be checked, it is presumed that the method to be checked
will present no bias.
All reference samples in this study were taken respecting the minimum sample masses as per Gy’s Theory of
Sampling. Following the company’s procedure to minimise the nugget effect, standard 50-g fire assays were
performed on triplicate samples whenever there was enough mass. Results showed equivalent particle size
distributions for the stopped belt and the cross-belt samples, and the bias test proved that the cross-belt
samples are unbiased. Therefore, it can be stated that the modified cross-belt sampler is appropriate to
estimate Jacobina’s plant feed grades. However, caution should be taken whenever the average feed grade
is considerably higher or lower than 2.5 g/t.
Therefore, considering both the studies carried out for low grade gold ores (Chieregati et al., 2017) and high
grade gold ores, the modified cutter design may reverse the bad reputation that cross-belt samplers have
acquired over the years.

Chieregati, A C, Amaral, Jr., E A and Souza, J C O, 2017. Validation of a modified cross-belt sampler for reconciliation
purposes. In: World Conference on Sampling and Blending, VIII. Perth, pp 247-251 (The Australasian Institute of
Mining and Metallurgy: Carlton).
Gy, P M, 1992. Sampling of heterogeneous and dynamic material systems: theories of heterogeneity, sampling and
homogenizing, 653 p (Elsevier: Amsterdam).
International Organization for Standardization, 1991. ISO 10226:1991 – Aluminium ores – Experimental methods for
checking the bias of sampling, September 1991.
Lyman, G, Nel, M, Lombard, F, Steinhaus, R and Bartlett, H, 2010. Bias testing of cross-belt samplers. In: Journal of the
Southern African Institute of Mining and Metallurgy, volume 110, number 6, pp 289-298 (The Southern African
Institute of Mining and Metallurgy: Johannesburg).
Pitard, F F, 1993. Pierre Gy’s sampling theory and sampling practice: heterogeneity, sampling correctness, and statistical
process control, second edition, 488 p (CRC Press: Boca Raton).
Robinson, G K, Sinnott, M D and Cleary, P W, 2010. Summary of results of ACARP project on cross-belt cutters. In: Journal
of the Southern African Institute of Mining and Metallurgy, volume 110, number 6, pp 331-338 (The Southern African
Institute of Mining and Metallurgy: Johannesburg).
Standards Australia, 2003. AS 2806.6-2003 – Aluminium ores – Sampling – Part 6: Methods for checking the bias of
sampling. Originated as AS 2806.6-1994, second edition, June 2003.


TABLE 3 – Value of t at 5% significance level (one-sided t-test).

Number of paired
data sets
20 1.729
21 1.725
22 1.721
23 1.717
24 1.714
25 1.711
26 1.708
27 1.706
28 1.703
29 1.701
30 1.699
31 1.697
32 1.696
33 1.694
34 1.692
35 1.691
36 1.690
37 1.688
38 1.687
39 1.686
40 1.685
41 1.684
42 1.683
43 1.682
44 1.681
45 1.680
46 1.679
47 1.679
48 1.678
49 1.677
50 1.677
51 1.676
61 1.671
81 1.664
121 1.658
241 1.651
 1.645

TABLE 4 – Required number of data sets, nr, determined by the value of the standardised difference, D.

Range of Required
standardized number of
difference data sets

D nr
0.30 ≤ D < 0.35 122
0.35 ≤ D < 0.40 90
0.40 ≤ D < 0.45 70
0.45 ≤ D < 0.50 55
0.50 ≤ D < 0.55 45
0.55 ≤ D < 0.60 38
0.60 ≤ D < 0.65 32
0.65 ≤ D < 0.70 28
0.70 ≤ D < 0.75 24
0.75 ≤ D < 0.80 21
0.80 ≤ D < 0.85 19
0.85 ≤ D < 0.90 17
0.90 ≤ D < 0.95 15
0.95 ≤ D < 1.00 14
1.00 ≤ D < 1.10 13
1.1 ≤ D < 1.2 11
1.2 ≤ D < 1.3 10
1.3 ≤ D < 1.4 8
1.4 ≤ D < 1.5 8
1.5 ≤ D < 1.6 7
1.6 ≤ D < 1.7 6
1.7 ≤ D < 1.8 6
1.8 ≤ D < 1.9 6
1.9 ≤ D < 2.0 5
2.0 ≤ D 5