Professional Documents
Culture Documents
ABSTRACT
During the development of a mining project, samples are taken for metallurgical tests. The sample
can be a composite sample made of material coming from a selection of various drilling products
or a sample coming from a mining test. This sample can represent several tonnes of material which
are analysed and split to be used in various tests.
In the current case, a composite sample is dedicated to semi-autogenous grinding (SAG) pilot
plant tests. The property of interest is then the size distribution of the material. The sample is
mixed and a subsample is taken to measure the size distribution which is considered as the feed
size distribution. Then the remaining material of the sample is divided in several subsamples for
batch grinding tests at different operating conditions. The product of each grinding test is then
analysed to measure its size distribution.
The grinding performances are analysed at the light of the measured size distributions for feed
and product. But what is the accuracy of the measurement of these size distributions? What is the
variability of the feed knowing each feed is just a subsample of the initial sample?
The Pierre Gy theory of sampling answers these questions.
The current paper presents, through a real example, the calculation of the measurement error of
the size distributions and the variability of the feed between the different tests. The uncertainties
of the grinding performances are then calculated and their impact on the results interpretation is
then analysed.
INTRODUCTION
During the development of a mining project, samples are distribution which is considered as the feed size distribution.
taken for metallurgical tests. The sample can be a composite Then the remaining material of the sample is divided in
sample made of material coming from a selection of various several subsamples which are used as feed of the grinding
drilling products or a sample coming from a mining test. This tests at different operating conditions. The characterisation
sample can represent several tonnes of material which are made on the first subsample is used to characterise all the feed
analysed and split to be used in various tests. In order to take subsamples. That is to say, for example, the same feed size
into account the ore variability in terms of hardness and size distribution will be used to calculate the performances of the
different tests even though a variability of the feeds occurs due
distribution, it is recommended to perform variability studies
to the division stage for obtaining them. The product of each
(Mosher and Bigg, 2002; Starkey, 2003; Meadows, Scinto and
grinding test is then analysed to measure its size distribution.
Starkey, 2011) using samples representing various horizons
or the years of production. But, for a given ore type, it can As the grinding performances are analysed at the light of
be necessary to perform many tests with different operating the measured size distributions for feed and product, any
conditions. In that case, the variability of the samples is not uncertainty in these measurements generates uncertainty in
due to the heterogeneity of distribution of the deposit, but the performance parameters of the tests and then uncertainty
only on the heterogeneity of constitution of the primary about the interpretation of results which can be done. Two
sample. questions arise:
In the current case, a metallurgical test campaign is dedicated 1. What is the accuracy of the measurement of the size
to the size reduction stage by semi-autogenous grinding distributions?
(SAG). A pilot plant scale SAG mill is used for several batch 2. What is the variability of the feed knowing each feed is
tests in various operating conditions. A composite sample just a subsample of the initial sample?
has been taken and constituted on field for this campaign. The Pierre Gy (1979) theory of sampling answers
The property of interest is then the size distribution of the these questions. It includes also the question about the
material. The sample is mixed and a subsample is taken representativeness of the composite sample. Indeed, in the
for various analyses including the measurement of the size frame of the development of a mining project, the pertinence
1. Scientific Manager, Caspeo, 3 Avenue Claude Guillemin, BP36009 45060 Orléans CEDEX 2, France. Email: s.brochot@caspeo.net
The moments of the probability distribution are used to in families is called ‘heterogeneity model’. It can be a fancy
characterise the measurement error. The first moment, the of the mind, however based on realistic assumptions. It can
mean, gives an evaluation of the bias, a systematic deviation also be obtained from specific experiments characterising
between the measurements and the true value. It measures the particles in terms of their heterogeneity. Such a model
the accuracy of the measurement. The second moment, the of heterogeneity has to be developed for each stage of the
variance, evaluates the reproducibility (or precision) of the sampling plan. Indeed, the material being classified before
measurement. subsampling, the heterogeneity changes in terms of size
distribution.
The overall measurement error (OE) includes a lot of
In the case of the size distribution, the critical component is
components which can be divided, following the Pierre Gy’s
one size class of particles and the critical content is that size
classification (Gy, 1979), into two main components: the total
class proportion. If the considered size class is entirely and
sampling error (TE) and the analytical error (AE).
solely represented by the particles of the family with index c,
The analytical error is due to the imperfection of the Equation 1 becomes:
protocols and devices used for analytical operations. When NF
(FE) = c 1 - 1 mfmc / miti p
1 - 2tc
concerning size distribution, the analysis is performed on the v2 + (2)
primary sample. The evaluation of the analysis error needs Ms ML tc
i=1
the decomposition of the protocols and procedures to find all
sources of error. Calculation rules and metrological approach Indeed, the content in the lot αL is the mass proportion of
are used to calculate the total analytical error. the considered family tc and the content αi is null for all the
families except for the considered family for which it is one.
The total sampling error has to take into account the The particles of the family c have a size in the range [xc 1; xc [
succession of bulk reductions which can occur during the +
and the particles of the other families have a size outside this
analysis protocol. It is then the sum of the total sampling range. Equation 2 gives the relative variance of the fundamental
errors at each stage (TEn). The sample preparation operations sampling error for the measurement of the proportion tc of the
generate the preparation error (PE) due to contamination, size class.
loss, chemical or physical alteration, unintentional or For the same reasons, if the considered size class is entirely
intentional mistakes. The operation of taking a small amount and solely composed of the particles of a set of families with
of material in a lot in order to obtain a sample generates the indexes in the subset Ic⊂{1…NF}, Equation 1 becomes:
sampling error in the proper sense (SE). Its components can
NF
v 2 (FE) = c 1 - 1 m f / miti 1 - 2f2 c + / miti p
be classified into continuous selection error (CE) resulting (3)
in the immaterial selection of the increments and increment Ms ML fc
idI i=1
materialisation error (ME) due to the particulate character c
of the material. The components of the continuous selection The particles of the families for which iIc have a size in the
error are due to the heterogeneity of the sampled material. range [xc 1; xc [ and the particles of the other families have a
+
The fundamental sampling error (FE) and the grouping and size outside this range. Equation 3 gives the relative variance
segregation error (GE), all together called short-range quality of the fundamental sampling error for the measurement of
fluctuation error (QE1), result from the heterogeneity of the proportion, fc , of the size class, which is the sum of the
constitution, while the long-range (QE2) and periodic (QE3) proportions of the families constituting it:
quality fluctuation errors, and the weighting error (WE)
comes from the heterogeneity of distribution in the space or fc = /t i
i d Ic
in the time. The increment delimitation error (DE) and the
increment extraction error (EE) constitute the materialisation If the considered size class is too wide in size range, it can
error. be necessary to cut it into sub size classes with smaller size
ranges assuming a better homogeneity inside each family in
Fundamental sampling error of the size class terms of unit mass. Similarly, if the particles in one size class
have various densities or shapes involving a wide range of
proportions unit masses, it can be necessary to use many density or shape
The theory of sampling is able to calculate the fundamental families.
sampling error starting from the description of the It is common (Gy, 1979) to use size and density classes for
heterogeneity of the material regarding the parameter the families used in Equations 1 to 3. Indeed, the size and
to measure: the particle size. The detailed description of density properties of the particles can easily be obtained
the heterogeneity is deduced from the size and density experimentally. The shape of the particles can also be taken
distribution analysis. The relative variance of the fundamental into account even though it is more difficult to describe
sampling error for the measurement of the content αL of the and estimate it. The particles of one family have then a size
critical component in the lot is given by Equation 1: in the range [ya + 1; ya [ , a density in the range [t b; t b+1[
N and a shape characterised by a shape factor f. The size and
F a -a 2
v 2 (FE) = c 1 - 1 m / miti e i L o (1) density ranges have to be as thin as possible to minimise the
Ms ML aL heterogeneity inside the family. In that case, the particles can
i=1
be considered as having a mean size xi and a mean density
In this formula, Ms is the mass of sample and ML the mass of t . The unit mass can then be calculated:
the lot. The particles are classified in different heterogeneity i
families according to their unit mass, ie the mass of one 3
mi = f xi ti
particle, and their content of critical component. The NF
families have to be as homogenous as possible, meaning that
all the particles in one family have more or less the same unit The sieve series {yα} used for the heterogeneity model can
masses, mi, and the same critical component content, αi. ti is have much more elements than the one {xi} used for analysis,
the mass proportion of the family in the lot. This classification trying to have {xi}⊂{yα}.
PILOT PLANT CAMPAIGN OF GRINDING TESTS sample of the -0.1 mm fraction is taken and analysed by
laser diffraction for the fine size distribution.
The metallurgical tests have been performed on a lateritic
• Each size fraction is crushed down to -1 mm if necessary
ore which is a blend of limonite, earthy saprolite and rocky
and divided using a sample splitter. Subsamples of the
saprolite facies. Limonite and earthy saprolite are characterised
size fractions are assayed by XRF.
by a large proportion of -100 μm, earthy saprolite having a
• The three subsamples obtained from the step 5 of the
larger proportion of millimetric particles. Rocky saprolite is
preparation procedure of the primary sample are assayed
characterised by a large proportion of +10 mm.
by XRF.
The metallurgical test campaign is focused on the size The equivalent dry mass of the lot is then ML = 4453 kg.
reduction stage by semi-autogenous grinding (SAG). A The mass of sample for particle size analysis is MS = 137 kg.
pilot plant scale SAG mill is used for several batch tests in The masses of material recovered in each size fraction are
various operating conditions. A primary sample of 5615 kg given in the Table 1 as well as the size class proportions and
(wet mass) coming from the deposit is used to perform these their associated relative fundamental sampling error (RFSE)
tests. The representativeness of this sample relatively to the calculated as two times the standard deviation (with 95 per
deposit is not discussed in this paper. The primary sample cent confidence) obtained from Equation 2.
is then considered as the lot and only the subsampling for
characterisation and the feed variability between the tests are The large fundamental sampling error for the measurement
studied here. of the 100 - 200 mm size proportion is due to the small number
of particles in the sample size fraction, three or four, in this
The primary sample is prepared following the procedure: class. In that case, the approximation done to obtain the
1. unloading and homogenisation with a front-end loader Equation 2 is not valid. It is necessary to return to the primary
2. division by coning and quartering to obtain a subsample formulae based on Poisson distribution of particle numbers
of 352 kg (wet mass) (Gy, 1979; Lyman et al, 2010). It can be the case also for the two
3. drying of the subsample in oven for moisture content following size classes for which the number of particles in the
measurement: 279 kg (dry mass) – 20.7 per cent moisture sample size fractions is <1000. The large error, 16.2 per cent,
for the smallest size classes is a consequence of the influence
4. division of the dried subsample to obtain 137 kg for
of the coarsest size classes in their representativeness.
sample characterisation
5. crushing down to -1 mm of the remaining part of the step 4 The two last columns of the Table 1 give the proportions of
and division by a sample splitter to obtain three samples the cumulative passing size distribution and their associated
for head assay relative fundamental sampling error. The d80 calculated by
log-linear interpolation is 90 mm. The associated relative
6. division by coning and quartering of the remaining part
fundamental sampling error calculated with Equation 5 is
of the step 2 to obtain a set of subsamples for the grinding
15 per cent or 14 mm as absolute error.
tests.
The products of the grinding tests are characterised in order Variability of test feeds
to calculate the grinding performances. The material used for each grinding test is a subsample of the
primary sample obtained from the step 6 of the preparation
Feed characterisation procedure of the primary sample. As any sampling process,
The primary sample is characterised through a subsample it is subjected to variability. Once again, the fundamental
issued from the step 4 of the preparation procedure of the sampling error gives the amplitude of this variability. The
primary sample. The results of this analysis are used as feed mean size distribution of the test feeds is supposed to be
characterisation for all the grinding tests performed with the the one measured from the characterisation subsample as
primary sample material. The analyses performed on the described above.
subsample of 137 kg are: The masses of material fed in the SAG mill vary from 58 kg
• Size distribution analysis by sieving with screen opening to 146 kg. The Table 2 gives the size class proportions and
100 mm, 50 mm, 11.2 mm, 5 mm, 2.8 mm, 1.4 mm and their associated relative fundamental sampling error (with
0.1 mm. The larger particles are less than 200 mm. A 95 per cent confidence) for these two extremes.
TABLE 1
Sample size distribution.
Size class ranges Weights (g) Proportions (%) Relative fundamental Cumulative passing Relative fundamental
(mm) sampling error of proportions (%) sampling error of
proportions (%) cumulative passing
proportions (%)
100 - 200 17 650 12.87 90.9 100.00 0.0
50 - 100 63 340 46.18 17.0 87.13 13.5
11.2 - 50 15 760 11.49 18.4 40.95 16.2
5 - 11.2 1399 1.02 16.8 29.46 16.2
2.8 - 5 798 0.58 16.3 28.44 16.2
1.4 - 2.8 676 0.49 16.2 27.86 16.2
0.1 - 1.4 3632 2.65 16.2 27.36 16.2
-0.1 33 900 24.72 16.2 24.72 16.2
Total 137 155 100.00
TABLE 2
Test feed size distribution variability.
Size class ranges (mm) Proportions (%) Relative fundamental sampling Relative fundamental sampling
error of proportions (%) 58 kg of error of proportions (%) 146 kg
feed of feed
100 - 200 12.87 141.0 88.0
50 - 100 46.18 26.4 16.5
11.2 - 50 11.49 28.5 17.8
5 - 11.2 1.02 26.0 16.3
2.8 - 5 0.58 25.2 15.8
1.4 - 2.8 0.49 25.1 15.7
0.1 - 1.4 2.65 25.0 15.6
-0.1 24.72 25.0 15.6
Here also, the large fundamental sampling error for the sampling error. The overall measurement error is limited
100 - 200 mm size proportion is due to the small number of to the analytical error. The d80 of the grinding test products
particles in this feed size fraction, one or two for 58 kg and varies from 24 mm to 77 mm and one test gives a product d80
three or four for 146 kg. The large error, 25.0 per cent for 58 kg of 92 mm. This apparently abnormal result giving a product
and 15.6 per cent for 146 kg, for the smallest size classes is also coarser than the feed (d80 of 90 mm) can be explained by the
a consequence of the influence of the coarsest size classes in large uncertainty of the feed d80. Indeed, the concerned test
their variability. was performed with only 71 kg which give an uncertainty
The uncertainty of the size distribution of the test feed due to of 25.8 per cent that is to say a feed d80 which can vary from
its overall measurement error adds to the intrinsic variability 66 mm to 114 mm. Considering a five per cent analytical error
of the feed as it is itself a sample. The sampling process for for the measurement of the product size distribution, there is
the characterisation and the sampling process for the test feed a seven per cent error for the product d80 which can vary from
are not really independent as they are taken in the same lot 85 mm to 99 mm.
without replacement. Nevertheless, the variance of the global
uncertainty of the feed size distribution can be approached Consequence regarding interpretation of
by the sum of the variance of the overall measurement error results
of the size distribution and the variance of the fundamental The test giving the coarsest product with a d80 of 77 mm
sampling error for test sampling. has been performed with a sample of 72 kg. The calculated
The Table 3 gives the proportions of the cumulative passing uncertainty on the feed d80 is then 25.6 per cent. The reduction
size distribution and their associated uncertainty for the ratio is 1.17. Taking into account the feed uncertainty and the
extremes of feed mass. The variance of the uncertainty is analytical error of the product size distribution, the error of
the sum of the variance of the fundamental sampling error the reduction ratio is 26.5 per cent.
for characterisation and of the variance of the fundamental The test giving the finer product with a d80 of 24 mm has been
sampling error for test sampling. The analytical error has performed with a sample of 79 kg. The calculated uncertainty
not been taken into account. The uncertainty of the feed d80 on the feed d80 is then 24.9 per cent. The reduction ratio is 3.74
calculated with the Equation 5 is 27.6 per cent for 58 kg and with an error of 25.8 per cent.
20.8 per cent for 146 kg (respectively 25 mm and 19 mm as
Another way to measure the grinding performances is to
absolute error).
calculate the operating work index:
Product characterisation W
Wi =
Once the grinding test performed, the mill product is entirely 10 - 10
analysed for size distribution. Consequently, there is no dP80 dF80
TABLE 3
Test feed cumulative passing size distribution uncertainty.
Size class ranges (mm) Cumulative passing proportions Uncertainty of cumulative passing Uncertainty of cumulative passing
(%) proportions (%) 58 kg of feed proportions (%) 146 kg of feed
100 - 200 100.00 0.0 0.0
50 - 100 87.13 24.8 18.7
11.2 - 50 40.95 29.9 22.5
5 - 11.2 29.46 29.8 22.5
2.8 - 5 28.44 29.8 22.5
1.4 - 2.8 27.86 29.8 22.5
0.1 - 1.4 27.36 29.8 22.5
-0.1 24.72 29.8 22.5
where W is the specific energy consumed during the test, dF80 and enrichment processes following the grinding stage are
and dP80 are respectively the feed and product d80. Considering impacted.
the specific energy as perfectly known (which is far from the
reality but the objective here is to know the influence of the CONCLUSIONS
size distribution uncertainty on the operating work index
The main result of this sampling approach to estimate the
estimate), the uncertainty relative variance on the operation
validity of metallurgical test works is the large uncertainty
work index, calculated using the rule of error propagation (for
at the level of the primary sample characterisation as well
the large errors used here, this rule start to be inapplicable),
as the variability of the feed of the different tests. And this
is given by:
uncertainty does not account the uncertainty in the primary
d v 2 (dP80) + dP80 v 2^dF 80h sample representativeness. Even though this observation
v 2 (wi) = F 80 seems pessimistic it has not to question the results of the
4 ` dF 80 - dP80 j
2
conducted metallurgical tests. Indeed, the specificity of
For the test giving the coarsest product, the calculated error the ore has well been taken into account. The results of this
on the operating work index is 166 per cent while it is 15.2 per first campaign have been confirmed by another one, at least
cent for the test giving the finest product. concerning the behaviour of the ore during grinding. But
all the results have to be read at the light of the mentioned
Four tests have been conducted with different per cent- uncertainty. That is to say, if some relationships are observed
solids, 25 per cent, 30 per cent, 40 per cent and 50 per cent, between performance parameters and operating conditions,
the other operating conditions being constant. Table 4 gives, they have to be compared to the expected ones. If the
for each of these four tests, the mass of sample used, the d80 differences are compatible with the level of uncertainty, the
of the product, the reduction ratio and its associated error, simplest relationships have to be chosen. Conversely, if the
and the operating work index and its associated error. The observed behaviour is far from the expected one or if the
Figure 1 displays the variation of the reduction ratio and of relationships do not appear clearly, the uncertainty analysis
the operating work index with the per cent-solids. The error has to discriminate between a bad test (issue during operation
bars correspond to one standard deviation (half error). or analysis), a specific behaviour of the ore, or a result due to
40 1.8
variability.
Operating work index Nevertheless, concerning the pilot plant test campaign
Reduction ratio described here, the following improvements can be mentioned
30 1.6 for recommendations:
Operating work index (kWh/t)
TABLE 4
Comparison of test results for various per cent-solids.
Solid content (%) Sample mass (kg) Product d80 (mm) Reduction ratio Reduction ratio Operating work Operating work
error (%) index (kWh/t) index error (%)
25 58 63 1.43 28.5 21.8 74.2
30 74 61 1.47 26.3 15.2 62.5
40 97 70 1.28 24.3 19.5 92.5
50 126 81 1.11 22.7 35.9 214.4
The approach presented here for the grinding tests can be Lyman, G, Nel, M, Lombard, F, Steinhaus, R and Bartlett, H, 2010.
applied for any kind of metallurgical test. The same question Bias testing of cross-belt samplers, The Journal of The Southern
of representativeness can be done for laboratory test such as African Institute of Mining and Metallurgy, 110:289-298.
the determination of the Bond work index or of the breakage Meadows, D, Scinto, P A and Starkey, J H, 2011. Seeking consensus
matrix. – How many samples and what testwork is required for a low
risk SAG circuit design, in Proceedings International Conference on
The results of such metallurgical test works are used Autogenous Grinding, Semiautogenous Grinding and High Pressure
to design and scale up the future processing plant. The Grinding Roll Technology 2011 (eds: K Major, B C Flintoff, B Klein
uncertainty at the pilot plant scale has to be propagated to and K McLeod) (Canadian Institute of Mining, Metallurgy and
the full-scale calculation and introduced in the financial risk. Petroleum: Vancouver).
Such an uncertainty analysis, conducted during prefeasibility Mosher, J and Bigg, T, 2002. Bench-scale and pilot plant tests for
or feasibility study, also helps in the definition of the future comminution circuit design, in Mineral Processing Plant Design,
metallurgical tests to reduce the uncertainty and to answer Practice, and Control (eds: A L Mular, D N Halbe and D J Barratt)
some questioning. This approach can also be done before the pp 123-135 (Society for Mining, Metallurgy, and Exploration, Inc:
first metallurgical tests. It is a way to have a better choice in Littleton).
the operating conditions to test and to select the more relevant Nageswararao, K, 1999. Reduced efficiency curves of industrial
ones. Indeed, it is always preferable to conduct less tests but hydrocyclones – An analysis for plant practice, Minerals
with a better accuracy. Engineering, 12:517-544.
Starkey, J H, 2003. Accurate, economical grinding circuit design using
REFERENCES SPI and Bond, in Proceedings XXII International Mineral Processing
Gy, P, 1979. Sampling of Particulate Materials, Theory and Practice, 431 p Congress (eds: L Lorenzen and D J Bradshaw) (The Southern
(Elsevier: Amsterdam). African Institute of Mining and Metallurgy: Johannesburg).