Professional Documents
Culture Documents
Real-Time Adaptive Estimation Framework For P80 in Hydrocyclones Overflow
Real-Time Adaptive Estimation Framework For P80 in Hydrocyclones Overflow
Real-Time Adaptive Estimation Framework For P80 in Hydrocyclones Overflow
hydrocyclones overflow
Benjamín Olivares*, Daniel Silva & Luis Yacher
SCAN Group, Contac Ingenieros, Santiago, Chile, Project Engineer, +56-2-22700900,
bolivares@contac.cl
Automation Department/Gerencia Plantas y Relaves, Minera Los Pelambres, Choapa, Chile, Senior
Control Engineer, +56-2-27983609, dsilva@pelambres.cl
Contac Ingenieros, Santiago, Chile, Manager, +56-2-22700900, lyacher@contac.cl
ABSTRACT
This paper presents the design and implementation of an on-line adaptive framework capable of
estimating the hydrocyclones overflow P80 coming from the ball mill by using nonlinear modelling
techniques in combination with the phenomenology of the process.
Determining the P80 is an important issue that has been widely discussed in numerous papers
related to mineral processing. And the reason is clear: there are significant advantages of
maintaining a consistent particle size with minimal variations under different operating conditions,
such as a better ball mill grinding and classification performance and an improved downstream
mineral recovery. However, the operating condition within a hydrocyclone is a complex process
and therefore, difficult to control. Moreover, the ideal conditions for applying the many modelling
and control techniques available in the literature or in the industry are not always technically or
economically feasible.
In this sense, the proposed estimation framework presents a solution to the aforementioned
problem via the design of a P80 nonlinear model based on empirical and phenomenological
knowledge and the implementation of a sub-optimal Bayesian filtering technique (more
specifically, particle filters), given their capability to combine real-time information measurements
from the process and empirical nonlinear models. Thus, it presents an improvement over the
traditional static-model approach as it takes its most important features while adding on-line
process characteristics and adaptive behaviour. This makes it particularly suitable for
implementing ball milling APC (advanced process control) strategies.
The efficacy of the proposed approach has been tested and put to work using real data from the
grinding plant at Minera Los Pelambres under steady state conditions and previous verification of
input data availability. These assumptions are necessary for the proper performance of the
approach.
The results validate that the estimation framework satisfactorily accomplishes the characterization
of the particles size variability (P80) in the hydrocyclones overflow, which in turn, readily allows
for deploying on-line control and process strategies for the appropriate selection, control and
optimization of the most relevant variables that are driving the efficiency of the particle separation
process.
–1–
INTRODUCTION
The purpose of size reduction in mineral processing grinding/classification is to achieve an
optimum release of mineralogical species obtained under current mining conditions. An efficient
operation at this stage has a direct impact on the recovery performance of the overall concentration
process and consequently, on the profitability of the mining business.
In order to achieve this goal, the best granulometric distribution which is associated with the
mineral to be recovered, must be obtained as a result of grinding and classification processes. This
is commonly represented by the passing 80% of the product distribution, namely, P80. Maintaining
this value with minimum variability is a complex task due to the high level of fluctation associated
to the process. Therefore, it becomes necessary to implement advanced process control strategies
(particularly, ball milling APC) to minimize disturbances in the system and achieve a particle size
within the desired range (Silva, 2012). From this point of view and to succeed in this task, P80 must
be monitored continuously with a high degree of reliability and availability.
This paper presents a solution for the problem described above through the design and
implementation of a framework for adaptive estimation of the P80 in the hydrocyclones overflow,
using a nonlinear structure combined with empirical and phenomenological knowledge of the
grinding and classification process. Furthermore, the proposed approach seeks to incorporate real-
time information of the process, adapting its estimation through the implementation of a particle
filtering-based estimation framework. Among the main contributions of this work are: (i) the
identification of critical variables influencing the process of obtaining a desired P80; (ii) generation
of a nonlinear model for estimation; and (iii) implementation of an adaptive estimation framework
capable of incorporating the process variability in real time.
The following section of this work presents the theoretical background on the problem of P80
estimation in hidrocyclone overflow and real-time adaptive framework based on particle filters.
The third section focuses on modelling aspects required for the incorporation of phenomenological
knowledge of the grinding process on the estimation framework. The fourth section shows the
construction of the P80 adaptive estimation framework based on particle filters. The fifth section
shows the experimental results of the proposed approach using data from a real grinding plant.
And finally, the sixth section presents the main conclusions and future work of this research.
THEORETICAL BACKGROUND
–2–
et al., 2004) summarizes and reviews the most common methods used for this task. Among them,
the determination of the hydrocyclone operation state by means of X-Ray, electrical impedance
tomography and accoustic emission sensing. In addition, there are estimation methods based on
empiric and/or semi-empiric modelling capable of providing relevant information of non-
observables or difficult to obtain measurements. These methods provide real time measurements at
relatively low maintenance costs, as well as being easily modified, which make them attractive for
measuring particle sizes at the hydrocyclone overflow. There are many types of methods that are
used to accomplish this task. For example, soft sensors based on fuzzy logic and artificial neural
networks for measuring the D50 cut-size were proposed by (Eren et al., 1997) and (Wong et al.,
2004,) respectively. Likewise, regression techniques as in (Nieto et al., 2009) are also presented,
where a virtual sensor estimates the solids sizes using weighted least squares (WLS) model and also
in (Sbarbaro & del Villar, 2010), where stepwise regression has been applied. During the last few
years, nonlinear mapping techniques, such as support vector machines (Sun, Wang & Zhang, 2008),
have also been used. It is worth mentioning among the classical models successfully utilized in the
resolution of hydrocyclones design and modelling problems, the Krebs model for hydrocyclones
dimensioning and the models of Lynch and Rao and Plitt and Nageswararo for simulation and
optimization (Gutiérrez & Sepúlveda, 1986). However, many of these approaches present their own
disadvantages, such as: (i) experimental data over-fitting; (ii) results or parameters without any
physical sense; and (iii) lack of parameters adaptation in the P80 estimation to compensate for the
effect of faulty/wrong measurements or to incorporate changes in the grinding system operating
conditions.
In this sense, particle filters (PF) have proved to be useful when trying to represent uncertainty in
nonlinear filtering process (Olivares et al., 2013). The choice of sequential Monte Carlo methods,
a.k.a., particle filters, is due to its notable ability to combine available process measurements and
analytic/empirical models. Therefore and considering the significant contribution of particle filter
algorithms to the implementation of filtering frameworks, it is deemed necessary to present a
summary of the main aspects associated to the formulation of particle-filtering-based estimation
models, which follows next.
k ( xk ) w k ( x k x k );
N (i) (i)
(1)
i 1
N
–3–
(i) (i) (i)
p ( y k | x k ) p ( x k | x k 1 )
w ( x k ) w k 1 i {1, ..., N }
(i) (i)
(i) (i)
(2)
q k ( x k | x k 1 )
where q k ( x k ) denotes the importance sampling density function. The most basic sequential Monte
Carlo implementation – the sequential importance sampling (SIS) particle filter – computes the
value of the particle weights w 0:( it) , by setting the importance density function equal to the a priori
state transition PDF p ( x t | x t 1 ) ; i.e., q t ( x 0:t | x 0:t 1 ) p ( x t | x t 1 ) . In that way, the weights for the newly
generated particles are evaluated from the likelihood of new observations. The efficiency of the
procedure improves as the variance of the importance weights is minimized. Further detail about
this powerful technique can be found in (Orchard & Vachtsevanos, 2009).
P 80 P 80 base (3)
where P 8 0 b a se is measured in microns and as mentioned, its resulting value will depend on
specific and particular plant design criteria.
However, the average grinding sizes can vary depending on the characteristics of the comminution
equipment and mineral properties, affecting thus, the classification operation. For this reason and to
correctly characterize these changes, the introduction of correction factors is needed.
In this sense, it is well known the influence of the solids percentage in the particles classification
and size distribution of the hydrocyclone overflow. Another factor of great impact is the influence
of the pressure drop that affects both the energy requirements and equipment wear rates. In
summary, (Arterburn, 1982) proposes a structure of an standard hydrocyclone configuration to
achieve the desired particle size. Along this line, the present work seeks to incorporate other factors
by considering the classification performance as a result of primary and secondary grinding
processes.
By modelling the effect of primary grinding (especially SAG Mill performance) over the P80 using a
PLS algorithm (SCAN, 2013), it is possible to analyze its coefficients to determine and weight the
impact of its input variables, as shown in Figure (2).
From here, it is clear that the variables having a greater influence in the P80 are throughput, current
draw and the fines percentage (less than 1¼” ore feed size) that enters the SAG mill. Consequently,
two new factors are incorporated: on one hand, by combining the effect of throughput and the
current draw of the SAG mill, an adjustment coefficient (associated with the specific energy
consumption (SEC SAG) and in units of [Ah/ton]) is generated, capturing thus, the effect of the
–4–
grinding media in the size reduction process and, on the other hand, there is a significant influence
of the fine feed size to the SAG mill, as this has a direct effect on both the operational performance
and the slurry viscosity of the SAG mill and consequently, to the amount of throughput to the ball
mill.
0.5
0.4
0.3
0.2
Weigths
0.1
-0.1
-0.2
Throughput SAG mill level Weight Speed Current Water Fines SAG Feed Size F80 Noise Level
Figure 1 Influence of SAG Mill operating parameters in the P80 estimation using PLS modelling
Regarding the influence of the secondary grinding, the aforementioned variables (pressure and
solids percentage in the hydrocyclone feed) by being controlled ones, are in turn affected by other
variables such as pump box level, number of open hydrocyclones, pump speed and volumetric
flow. Nevertheless, the effect of the specific energy consumption in the ball mill (SEC BM) for size
reduction, as well as the effect of the circulating load ratio (CLR) are not mentioned, but they are
included in the P80 estimation model. Therefore, with the inclusion of the adjustment factors,
equation (3) takes the form of the following equation:
6
P 8 0 P 8 0 b a se C i ( x i ) (4)
i 1
x1 x2 x3 x4 x5 x6
In order to determine the effect of each factor on the size distribution and due to the nonlinear
relationship of the variables with the P80, a potential function is considered for its construction.
i
C i ( x i ) ( i x i ) i {1, ..., 6} (5)
–5–
where the parameter is a normalization value associated with a nominal operating point. The
second parameter, , indicates the direction and proportional magnitude of the variable to P80,
which can be determined empirically. Therefore, equation (4) can be rewritten as follows:
6
i
P 8 0 P 8 0 b a se ( i x i ) (6)
i 1
Once the base model to estimate the real-time P80 is defined, it is necessary to generate a
framework to incorporate information from the actual process and to adapt to noisy observations,
which are intrinsic to the process of comminution.
P 8 0 j ( k ) P 8 0 b a se C i ( i x i ( k )) ( k ) w j ( k )
i
i 1 (7)
( k 1) ( k ) ( k )
Measurement equation:
–6–
j
j (k ) P80 j (k )
P 8 0 la b ( k ) v(k ) (8)
j
j (k )
where P 8 0 j is a state representing P80 value at the output of each battery, in this case
j { M B 4, M B 5} given the configuration of line 1 (see figure (2)); is a non-observable state
associated with an unknown model parameter that is required to explain temporary differences
with the expected behavior (which are basically and specifically, attributed to the laboratory P80
sampling procedures); P 8 0 la b is the P80 measurement as calculated by the lab; w j , are non-
Gaussian process noises representing uncertainty on the a priori state estimates and v is measured
noise. parameter is a ratio representing the contribution of each battery in the final composite
slurry flow processed by line 1 milling/classification process, where the condition that j 1
n
must be met and whose values can be calculated by the ratio between the flow feed to each
hydrocyclone battery as shown below:
feed M B 4
MB 4 ;
feed M B 4 feed M B 5
(9)
feed M B 5
MB5 .
feed M B 4 feed M B 5
where fe e d M B 4 and fe e d M B 5 are feed flows of each hydrocyclone battery associated to MB4 and
MB5 ball mills, respectively. This approach allows providing continuous estimates not only when
both mills are operating, but also when only one mill is contributing and the other one is not
available (for whatever reasons).
Figure 2 Process flow diagram of the grinding and classification circuit, line 1
–7–
It should also be noted that, even in the absence of new lab measurements, the proposed scheme
continues estimating the values of P80 in the overflow of each hydrocyclone battery, which goes to
validate the availability and robustness of the proposed approach. However, and as expected, the
state value will not be updated and adjusted to new lab measurements and over time, will
increase the uncertainty in the estimates. Despite the above, it has been experimentally observed
that the estimates are still robust and valid as long as there are no drastic changes in the operating
conditions or in the ore characteristics.
350
Estimated P80 MB4
300
250
P80 [um]
200
150
100
0 100 200 300 400 500 600 700 800 900 1000
N° Observations
350
Estimated P80 MB5
300
250
P80 [um]
200
150
100
0 100 200 300 400 500 600 700 800 900 1000
N° Observations
Figure 3 P80 estimate signals associated to ball mill MB4 and MB5
In addition to the previous results, and once the P80 at the output of each hydrocyclone battery is
obtained, it is possible to back-calculate the composite estimated signal for the line and compare it
with the corresponding lab result, verifying thus empirically, the proposed framework
performance. See figure (4).
–8–
350
Observed P80
Estimated P80
300
250
P80 [um]
200
150
100
0 100 200 300 400 500 600 700 800 900 1000
N° Observations
Figure 4 Comparison of P80 observed signal and P80 estimated signal for a period from October 2013
to April 2014 (RMSE = 19.94)
There is a clear emerging pattern that proves that the classification process variability is indeed,
captured. Nevertheless, there are some differences, due to the noisy nature of the measurements
and the presence of other slurry inputs and outputs which are not accounted for in the model (due
to the absence of flow measurements coming from the other two grinding circuits, line 2 and line 3),
which should eventually, be corrected. In fact, the effect of phase differences was also studied
during the model implementation, but no better performance was observed. Analog results are
obtained for line 2 of Minera Los Pelambres grinding plant.
–9–
As a future work, it is considered to improve the performance of the estimation framework by
incorporating more phenomenological aspects in the construction of the estimation model for the
P80, such as the effects of mineral hardness and wear rates of major equipment (SAG mill, ball mill
and hydrocyclones).
REFERENCES
Arterburn, R. A. (1982). The sizing and selection of hydrocyclones. Design and Istallation of Comminution
Circuits, 1, 597-607.
Eren, H., Fung, C. C., Wong, K. W., & Gupta, A. (1997). Artifical neural networks in estimation of
hydrocyclone parameter D50C with unusual input variables. Instrumentation and Measurement, IEEE
Transactions on, 46(4), 908-912.
Gutiérrez, L., & Sepúlveda, J. (1986). Dimensionamiento y optimización de plantas concentradoreas mediante
técnicas de modelación matemática. CIMM.
Henríquez, F., Silva, D., & Jiménez, C. (2012). Improving ball mills operations by applying MPC & advanced
control systems at Minera Los Pelambres. XXVI International Mineral Processing Congress (IMPC 2012).
Janse van Vuuren, M. J. (2011). On-line monitoring of hydrocyclones by use of image analysis. Doctoral
dissertation, Stellenbosch: University of Stellenbosch.
Magne, L., & Titichoca, G. (2004). Conminución de Minerales. Universidad de Santiago, Departamento de
Ingeniería Metalúrgica.
Neesse, Th., Schneider, M., Goyk, V., & Tiefel, H. (2004). Measuring the operating state of the hydrocyclone.
Minerals Engineering, 17, 697-703.
Nieto, L., Olivares, J., Gatica, J., Ramos, B., & Olmos, H. (2009). A virtual sensor for estimating particle size of
hydrocyclones overflow. IFACMMM 2009.
Olivares, B., Cerda, M., Orchard, M., & Silva, J. (2013). Particle-filtering-based prognosis framework for energy
storage devices with a statistical characterization of state-of-health regeneration phenomena.
Instrumentation and Measurement, IEEE Transactions on, 62(2), 364-376.
Orchard, M., & Vachtsevanos, G. (2009). A particle filtering approach for on-line fault diagnosis and failure
prognosis. Transactions of the Institute of Measurement and Control, vol. 31, no. 3-4, pp. 221-246, June 2009.
Sbárbaro, D., & Del Villar, R. (2010). Advanced Control and Supervision of Mineral Processing Plants. Advances
in Industrial Control. Springer (p. 311).
SCAN (2013), software developed by the Chilean company CONTAC Ingenieros. http://www.contac.cl/scan.
Silva, D. (2012). State of the art & Best practices in Advanced Control at Minera Los Pelambres. Automining
2012.
Sun, Z., Wang, H., & Zhang, Z. (2008). Soft sensing of overflow particle size distribuctions in hydrocuclones
using a combined method. Tsinghua Science & Technology, 13(1), 47-53.
Wong, K. W., Ong, Y. S., Eren, H., & Fung, C. C. (2004). Hybrid fuzzy modelling using memetic algorithm for
hydrocyclone control. Proceedings of 2004 International Conference on Machine Learning and Cybernetics,
IEEE.
– 10 –