Minerals Engineering 53 (2013) 18

Contents lists available at ScienceDirect

Minerals Engineering
journal homepage: www.elsevier.com/locate/mineng

Remaining useful life prediction of grinding mill liners using an articial

neural network
Farzaneh Ahmadzadeh , Jan Lundberg
Lule University of Technology, Division of Operation and Maintenance, Lule, Sweden

a r t i c l e

i n f o

Article history:
Received 6 April 2013
Accepted 30 May 2013
Available online 5 August 2013
Keywords:
Remaining useful life
Articial neural network
Principal component analysis
Wear prediction
Maintenance scheduling
Mill liners

a b s t r a c t
Knowing the remaining useful life of grinding mill liners would greatly facilitate maintenance decisions.
Now, a mill must be stopped periodically so that the maintenance engineer can enter, measure the liners
wear, and make the appropriate maintenance decision. As mill stoppage leads to heavy production losses,
the main aim of this study is to develop a method which predicts the remaining useful life of the liners,
without needing to stop the mill. Because of the proven ability of articial neural networks (ANNs) to recognize complex relationships between input and output variables, as well as its adaptive and parallel
information-processing structure, an ANN has been designed based on the various process parameters
which inuence wear of the liners. The process parameters were considered as inputs while remaining
height and remaining life of the liners were outputs. The results show remarkably high degree of correlation between the input and output variables. The performance of the neural network model is very consistent for data used for training (seen) and testing (unseen).
2013 Elsevier Ltd. All rights reserved.

1. Introduction
Todays industries are increasingly using condition based maintenance (CBM) to minimize breakdowns and their impact on performance, reduce maintenance intervals and consequent costs,
improve production efciency, and ensure safety. The development
of a maintenance system with intelligent features in fault detection
and knowledge accumulation for mechanical structures is a goal
for researchers; such a system would greatly assist industries, as
it is now almost impossible to manually analyze rapidly growing
data to extract valuable decision-making information. In recent
decades, a great deal of research has sought methods of predicting
or estimating the remaining useful life (RUL) of critical components in various industries. Knowing the RUL of an asset has an impact on planning maintenance activities, spare parts provision,
operational performance, and protability (Jardine et al., 2006; Altay and Green, 2006; Elwany and Gebraeel, 2008; Wang et al.,
2009; Kim and Kuo, 2009; Papakostas et al., 2010).
Ore grinding mills are heavy duty pieces of equipment that
work 24 h a day in highly abrasive environments. From an economic point of view, it is important to keep these mills in operation
and minimize the downtime for maintenance or repair, because a
drop in production caused by a both scheduled and unscheduled
stoppages lead to monetary losses. The auto-genus mill which is
used in mineral processing is important for particle size reduction
Corresponding author. Tel.: +46 920 49 2106.
E-mail address: Farzaneh.Ahmadzadeh@ltu.se (F. Ahmadzadeh).
0892-6875/$ - see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.mineng.2013.05.026

and high metal recovery. Among its most critical components are
the liners which protect the mills shell and are used to lift the
charge (ore) inside the mill, thus enhancing grinding performance,
Fig. 1. Because their wear inuences the grinding performance in
the context of metal recovery, the mill needs to stop occasionally
for the maintenance engineer to enter the mill and measure the
wear, but each stoppage leads to heavy production losses. The signicant impact of mill liners on the monetary return for the mill
owner has led to studies of maintenance activities performed on
mill liners, such as wear measurement, replacement and maintenance scheduling.
Although many studies have been carried out on the effect of
mill liners on grinding mill performance (Cleary, 2001; Santarisi
and Almomany, 2005; Kalala et al., 2008; Yahyaei et al., 2009; Dandotiya and Lundberg, 2012), the ability to predict or measure the
liners wear without stopping the mill has not been considered.
Therefore, the main focus of the present research is optimization
of wear measurement, as well as RUL predictions for optimal
replacement and maintenance scheduling to maximize both performance and prots.
A review of approaches to RUL prediction shows that an articial neural network (ANN) is a powerful tool; it can readily address
modeling problems that are analytically difcult and for which
conventional approaches are not practical, including complex
physical processes with nonlinear, high-order, and time-varying
dynamics and those for which analytic models do not yet exist.
Zhang and Ganesan (1997) used a self-organizing neural network
for multivariable trending of fault development to estimate the

F. Ahmadzadeh, J. Lundberg / Minerals Engineering 53 (2013) 18

operating information, particularly for the wear out phase of liner

measurement. The lack of accurate data for this part of the liner
could have hampered the study. However, data requirements were
fullled by selecting the shell feed (type1 of the lifter bar (LB1); see
Fig. 1b) of the liner from 2008 to 2011 for two life cycles.
2.1. Data sources
The following sources were used to gather condition monitoring
(CM) and process data:
 Metso mineral for wear measurement CM data.
 Boliden mineral for affecting factors (process data).
Utmost care was taken to ensure that the collected data were as
accurate as possible. There were no data for some periods; in other
periods, the grinding mill worked only a few hours per day, etc. We
sought to determine why these occurred.
Fig. 1. a Mill liners inside the mill, (b) shell feed lifter bars.

RUL of a bearing system. Shao and Nezu (2000) proposed a new

concept called a progression-based prediction of remaining life to
estimate the RUL of a bearing. This concept manipulated the variables determined from online measurements via a compound
model of a neural network. Byington et al. (2004) developed a neural network for remaining life predictions for aircraft actuator components. Yu et al. (2006) presented a neural network model to
predict behavior of a boring process during its full life cycle. Mazhar et al. (2007) estimated the remaining life of used components
in consumer products by using articial neural networks. Runqing
et al. (2007) dealt with residual life predictions for ball bearings
based on a self-organizing map and back propagation neural network methods. Similarly, Huang (2007) proposed a method to predict a ball bearings RUL based on a self-organizing map and back
propagation neural network methods. Gebraeel and Lawley
(2008) developed a neural network to estimate the RUL of rolling
element bearings by monitoring their vibrations. In this study,
ANN has been applied to predict the RUL of the liner with respect
to the remaining height and remaining life.
The paper is organized as follows: Section 2 describes the data
collection; Section 3 explains methodology; results and discussion
are provided in Section 4. Section 5 offers a conclusion.

2. Data collection
For the mill liners, the development of an RUL assessment based
on life cycle data analysis is hampered due to the unavailability of

2.1.1. Data from Metso mineral

For this study, an important source of raw data on liner maintenance, inspection and replacement schedule was Metso minerals.
The most difcult part was collecting data on liner wear because
there were fewer measurements. Because total value of lost production during any mill stoppage is extremely high, it is not economical to stop the mill at intervals and measure liner wear,
except for maintenance, inspection, installation, replacement.
Thus, data were collected over different life cycles (from one installation to the next replacement). The solid circles and triangle in
Fig. 2 show LB1 remaining height and remaining life during one life
cycle; other data are generated by an interpolating technique,
Piecewise Cubic Hermite Interpolating Polynomial (PCHIP) and
Spline. PCHIP is the best choice for this study, because it nds values of an underlying interpolating function at intermediate points.
It also preserves the shape of the data and respects monotonicity,
and these data have monotonically deceasing characteristics.
2.1.2. Data from Boliden mineral
Raw process data on the grinding mill, collected from the Boliden mineral databases, were treated to extract the information
used in the models. Information on the important inuencing factors for liner wear was collected in discussions with expert personnel and engineers from the mining company. These data include
the ore type, ore feed (tonne/h), power (kW), angular speed (% of
centrifugal critical speed), torque (% of the max torque), water
addition (m3/h), grinding energy (kW h/tonne), load (tonne). This
information comprised the practical complexities of the grinding
process inside the mill, including physical explanations of the

Fig. 2. Generated remaining height and remaining life data by PCHIP interpolating method.

F. Ahmadzadeh, J. Lundberg / Minerals Engineering 53 (2013) 18

energy consumption, power and speed uctuation, etc. during different wear phases of the mill liners.
3. Prediction of remaining useful life by articial neural
network
As every component operates under different conditions, the
actual operating life is dependent on the real conditions of use
and, therefore, the operating conditions must be taken into account
when determining the life of the product and/or its components. In
other words, the remaining useful life is a function of the actual life
(LA) under given operating conditions of use. But real conditions or
process data will affect the wear out process. The unique nature of
each data type, process and wear out data, requires specialized
analysis techniques. In this study, ANN has been employed to nd
nonlinear or linear relationships between input (process data) and
output (CM) variables.
3.1. Data preparation
As mentioned earlier, two important factors showing remaining
life, namely, height (mm) and volume (life (%)) of the shell feed
(LB1) were considered output variables (see Fig. 1b the remaining
volume is covered by darker color). Ore type, ore feed (tonne/h),
power (kW), angular speed (% of centrifugal critical speed), torque
(% of the max torque), water addition (m3/h), grinding energy
(kW h/tonne), load (tonne) were recorded over the life cycle as input variables.
Two life cycles of the lifter bar were combined, the rst from
2008-04-15 to 2009-09-23and the second from 2009-12-22 to
2011-04-10. We gathered 886 sets of data in both cycles. At rst
cycle 80% (347 sets) of the data were used for training and 20%
(87 sets) for testing the network. The testing data was grouped in
the multiples of 6, like 6, 12, and 18. of the total data in rst cycle.
3.2. ANN architecture
Articial neural networks are adaptive and have parallel information-processing structures with the ability to build functional
relationships between data and to provide powerful tools for nonlinear, multidimensional interpolations. This aspect of neural networks makes it possible to capture and interpret the existing
highly complex nonlinear relationships between input and output

parameters that are frequently poorly understood (Eskandari et al.,

2004). An ANN is a system consisting of processing elements (PE)
with links between them. A certain arrangement of the PEs and
links produce a certain ANN model, suitable for certain tasks.
A Multi-Layer Perceptron (MLP) is a kind of feed-forward ANN
model consisting of three adjacent layers; the input, hidden and
output layers. Each layer has several PEs. MLPs learn from input
output samples to become clever i.e. capable of giving outputs
based on inputs which it has not seen before. The learning process
employs a learning algorithm, during which the MLP develops a
mapping function between the inputs and outputs. Basically, in a
learning process, the input PEs receive data from the external environment and pass them to the hidden PEs, which are responsible
for simple, yet, useful mathematical computations involving the
weight of the links and the input values. The results from the hidden PEs are mapped onto appropriate threshold function of each PE
and the nal outputs are produced. The output values then become
inputs to all PEs in the adjacent layer (either the second hidden
layer or the output layer), and the computation processes are repeated throughout the layers until nally, output values are produced at the output PEs. At this stage, an output error value is
calculated by computing the difference between the MLPs and
the actual outputs.
The entire training process is iterative in nature, and stops
when an acceptably small error is achieved. At completion of a
learning process, the MLP should be able to give output solution
(s) for any given set of input data based on the generalized mapping that it has developed Junita and Brian (2008).
The architecture of the proposed three layer network consists of
two hidden layers of sigmoid transfer function followed by an output layer of a linear transfer function. Hidden layers with nonlinear
transfer functions allow the network to learn nonlinear and linear
relationships between input and output variables. The linear transfer function in the output layer lets the network produce outputs
outside the range [ 1, 1].
The number of inputs to the proposed network is given by the
number of available inputs or process data (ore type, ore feed,
power, speed, torque, water addition, grinding energy, load); the
number of neurons in the output layer is constrained to two, as
the output required two parameters, remaining height and remaining life, of the liner. The number and size of layers between network inputs and the output layer are determined by testing
several combinations of numbers of layers and various numbers

Fig. 3. Architecture of the proposed MLP Network.

F. Ahmadzadeh, J. Lundberg / Minerals Engineering 53 (2013) 18

of neurons in each layer. Each of the selected combinations is

tested with several different initial conditions to guarantee that
the proposed model is the best solution. The resulting network
consists of nine inputs, two hidden layers of 25 and 50 neurons
respectively, and two nodes at output layers. The architecture of
the proposed neural network is shown in Fig. 3.
3.3. Principal component analysis
Most of the real-world data samples used to train articial neural networks (ANNs) consist of correlated information from overlapping input instances. Correlation in sampled data normally
creates confusion during the learning process and, thus, degrades
the generalization capability. The performance of a MLP depends
on its generalization capability which, in turn, is dependent on
the data representation. A set of data presented to a MLP ought
not to consist of correlated information; correlated data reduce
the distinctiveness of data representation, introducing confusion
to the MLP model during the learning process and producing a
model with low generalization capability to resolve unseen data.
This suggests a need to eliminate correlation in the sample data before they are presented to a MLP. This can be achieved by applying
the principal component analysis (PCA) technique.
PCA involves a mathematical procedure that transforms a number of (possibly) correlated variables into a (smaller) number of
uncorrelated variables called principal components. Basically, Junita and Brian (2008) explained that the PCA technique consists of
nding linear transformations V=(v1, v2, v3, . . ., yp) of the original
components X=(x1, x2, x3,. . ., xQ) that have a property of being
uncorrelated. In other words, the V components are chosen in such
a way that v1 has maximum variance; v2 has maximum variance
subject to being uncorrelated with v1, and so forth. The rst step
in the PCA algorithm is to normalize the components so that they
have zero mean and unity variance. Then, an orthogonal method is
used to compute the principal components of the normalized components see also, Hong and Wu (2012) and Campbell and Atchley
(1981).
In this research, PCA was used to eliminate those components
contributing less than 5% to the total variation in the data set.
The results after applying the PCA indicated signicant redundancy
in the training data set, as the principal component analysis reduced the number of input vectors from 9 to 5 (see Fig. 4).
3.4. Training and performance testing
A problem that occurs during neural network training is overtting or over-training. The training style is the supervised learning in which the rules are supplemented with a set of examples

(the training set) of proper network behavior. The training set consists of inputs and the corresponding correct outputs (targets). One
of the most powerful learning algorithms, the LevenbergMarquardt algorithm (Haykin, 1999), has been used to train networks.
In function approximation problems, this algorithm is considered
to have the fastest convergence. Although the network remembers
the training examples, however, it does not learn to generalize to
new situations.
The MATLAB function trainbr, used to train the proposed network, has a built-in procedure, Bayesian regularization, designed
to overcome over-tting. This technique has been documented as
a better generalization procedure for function approximation problems. For the most efcient training, input data are pre-processed
before training. The selected training function trainbr works best
when the network inputs and targets are scaled to fall approximately in the range [ 1, 1].
In this study, pre-processing was done by using the function
prestd. After the training was completed, the network was tested
for its learning and generalization capabilities. The test for its
learning ability was conducted by testing its ability to produce outputs for the sets of inputs (seen data) used in the training phase.
The networks outputs had a correlation coefcient of about
0.99987 with the desired (actual) outputs. The test for the networks generalization ability was carried out by investigating its
ability to respond to the input sets (unseen data) not included in
the training process.
The results of the proposed neural network showed a remarkably high degree of correlation between the input and output variables. Clearly, the neural network model could handle the complex
nonlinear interrelationships between variables. Furthermore, the
performance of the neural network model was very consistent
for both training and test data. Finally, there was not a considerable difference in the networks output when trained with either
seen or unseen data; the proposed model accurately approximated
the inputoutput function.
4. Results and discussion
Remaining life estimation using nonlinear inputs is far more
complicated than using linear inputs, especially in the case of an
intricate mixture of uctuating and unpredictable trends. Note that
the results produced by the proposed neural network for remaining life assessment are associated with higher levels of certainty
because statistical analysis techniques were employed.
The best available functions and procedures were utilized to
pre-process the inputs, train the network and post-process the
outputs of the proposed model. The blue line (variation/noise) in
Fig. 5 shows the prediction of the ANN in remaining height and

Fig. 4. Applying the PCA and reduction of the inputs vector dimension from 9 to 5.

F. Ahmadzadeh, J. Lundberg / Minerals Engineering 53 (2013) 18

Fig. 5. Training capability: real and predicted height and remaining life proposed ANN for seen data (80% of the total data in rst cycle).

Fig. 6. Generalization capability: real and predicted remaining height and remaining life proposed ANN for unseen data (20% of total data in rst cycle).

F. Ahmadzadeh, J. Lundberg / Minerals Engineering 53 (2013) 18

Fig. 7. Prediction of ANN for remaining height and remaining life for two life cycles.

remaining life of the liners while the red line (steady variation)
shows real outputs. This shows the high ability of the network to
produce outputs for the 80% sets of inputs (347 data sets in the rst
cycle) used in the training phase. The gure shows a high correlation between actual and predicted output. For generalization ability, we investigated its ability to respond to 20% of the input sets
(87 data sets of unseen data for the rst cycle) which were not included in the training process; see Fig. 6. A high correlation between the networks predicted output and the actual output is
clearly visible. Note that the x-axis in Fig. 7 are showing the dates
when the measurements are performed for both cycles. Stops of
the mills are included and thus the x-axis are not showing only
the total running time of the mills. It can be discussed if the total
running time of the mill should be presented in the x-axis instead
of actual date when the measurements are performed. However,
from the maintenance point of view it is more practical to use the
date based presentation, since otherwise it is difcult to keep track
of the actual wear at specic time. But in Figs. 5 and 6 because of the
nature of selecting the training and testing data (80%:20%) the xaxis are just the data points. It is why (for example) in 87th data
point the height has come down to around 200 mm in Fig. 6 while
at the same data point in Fig. 5 it is around 350 mm. And it is because the input parameters in the training phase are not the same
as the inputs for test phase. Thus it can be concluded that the neural
network is capable of predicting of the remaining life of the liners.
Another test of the proposed networks ability to predict
remaining height and remaining life was training the designed
ANN with whole data sets belonging to the rst life cycle (434 data
sets) and testing its predicting ability for all data in the second cycle. The results in Fig. 7 show a remarkably high correlation between real and predicted outputs.
To measure the proposed networks performance, we calculated
the relative error for its learning and generalization capability. As

shown in Figs. 8ac, the maximum relative error was less than
6% and 4% for remaining height for seen and unseen data respectively. As shown in Fig. 8bd, it was less than 10% for the remaining
life for both seen and unseen data. The network predicted the
remaining height and remaining life of the liners with accuracy
greater than 90%. Therefore, the proposed model can approximate
the inputoutput function with high accuracy.

5. Conclusions
This study presents a comprehensive approach called ANN
which can provide remaining life estimates with higher levels of
certainty. Degradation and condition monitoring data were analyzed by developing a three layer feed-forward back-propagation
neural network, consisting of nine inputs which reduced to ve
after applying PCA, two hidden layers of 25 and 50 neurons respectively, and two nodes at output layers, remaining height (mm) and
remaining life (%). The results had 90% accuracy, and the performance of the neural network model was very consistent for both
training data and test data. This is a critical advance in the maintenance procedure for grinding mill liners; it allows better understanding of the mills service requirements and can predict the
remaining life of the liners while they remain in operation; it is
not necessary to stop the mill for every maintenance activity, preventing high monetary losses.
In summary, the main advantage of the proposed neural network is, this methodology does not require disassembly of the liners or stoppage of grinding mills to make decisions on replacement
or maintenance. In addition, it does not require expensive and
sophisticated equipment for data recording and analysis. Moreover, the ANN model can accommodate a wide range of input variables with complex and nonlinear input trends/patterns. However,

F. Ahmadzadeh, J. Lundberg / Minerals Engineering 53 (2013) 18

Fig. 8. Relative error for predicted remaining height and remaining life for training and testing phase (seen and unseen data).

it requires large amounts of data for training and tuning. Consequently, the dynamic nature of the proposed methodology opens
avenues for future study of life cycle data in different categories
of products.
Acknowledgements
The authors would like to thank Boliden, Metso Mineral for supporting this research and permission to publish this article. Special
appreciation is extended to the operating, maintenance engineers
Jan Burstedt (Boliden), Lars Furtenbach and Magnus Eriksson (Metso) for sharing their valuable experience, knowledge, and data to
improve the paper.
References
Altay, N., Green III, W.G., 2006. OR/MS research in disaster operations management.
European Journal of Operational Research 175, 475493.
Byington, C.S., Watson, M., Edwards, D., 2004. Data-driven neural network
methodology to remaining life predictions for aircraft actuator components.
IEEE, Piscataway, NJ, USA.
Campbell, N.A., Atchley, W.R., 1981. The geometry of canonical variate analysis.
Systematic Zoology 30, 268280.
Cleary, P.W., 2001. Charge behavior and power consumption in ball mills:
sensitivity to mill operating conditions, liner geometry and charge
composition. International Journal of Mineral Processing. 63, 79114.

Dandotiya, R., Lundberg, J., 2012. Economic model for maintenance decision: a case
study for mill liners. Journal of Quality in Maintenance Engineering 18, 79
97.
Elwany, A.H., Gebraeel, N.Z., 2008. Sensor-driven prognostic models for equipment
replacement and spare parts inventory. IIE Transactions 40, 629639.
Eskandari, H., Rezaee, M.R., Mohammadnia, M., 2004. Application of multiple
regression and articial neural network techniques to predict shear wave
velocity from wireline log data for a carbonate reservoir, South Iran. Canadian
Society of Exploration Geophysicists CSEG Recorder 29, 4248.
Gebraeel, N.Z., Lawley, M.A., 2008. A neural network degradation model for
computing and updating residual life distributions. IEEE Transactions on
Automation Science and Engineering 51, 154163.
Haykin, S., 1999. Neural Networks: A Comprehensive Foundation. Prentice Hall,
Upper Saddle River, New Jersey.
Hong, Y., Wu, C., 2012. Day-ahead electricity price forecasting using a hybrid
principal component analysis network. Energies 5, 47114725.
Huang, R., 2007. Residual life predictions for ball bearings based on self-organizing
map and back propagation neural network methods. Mechanical Systems and
Signal Processing 21 (1), 193207.
Jardine, A.K.S., Lin, D., Banjevic, D., 2006. A review on machinery diagnostics and
prognostics implementing condition-based maintenance. Mechanical Systems
and Signal Processing 20, 14831510.
Junita, M., Brian, S., 2008. Improved neural network performance using principal
component analysis on Matlab. International Journal of the Computer the
Internet and Management 16, 18.
Kalala, J.T., Breetzke, M., Moys, M.H., 2008. Study of the inuence of liner wear on
the load behavior of an industrial dry tumbling mill using the discrete element
method (DEM). International Journal of Mineral Processing 86, 3339.
Kim, K.O., Kuo, W., 2009. Optimal burn-in for maximizing reliability of repairable
non-series systems. European Journal of Operational Research 193, 140
151.

F. Ahmadzadeh, J. Lundberg / Minerals Engineering 53 (2013) 18

Mazhar, M.I., Kara, S., Kaebernick, H., 2007. Remaining life estimation of used
components in consumer products: life cycle data analysis by Weibull and
articial neural networks. Journal of Operations Management 25, 11841193.
Papakostas, N., Papachatzakis, P., Xanthakis, V., Mourtzis, D., Chryssolouris, G., 2010.
An approach to operational aircraft maintenance planning. Decision Support
Systems 48, 604612.
Runqing, H., Lifeng, X., Xinglin, L., Richard, L., Hai, Q., Jay, L., 2007. Residual life
predictions for ball bearings based on self-organizing map and back
propagation neural network methods. Mechanical Systems and Signal
Processing 211, 193207.
Santarisi, N.S., Almomany, R.M., 2005. Mathematical modeling of wear rate of nonrepairable parts and their replacement strategies: cement mill liners as a case
study. Journal of Quality in Maintenance Engineering 11, 6881.
Shao, Y., Nezu, K., 2000. Prognosis of remaining bearing life using neural networks.
proceedings of the institution of mechanical engineers. Part I: Journal of
Systems and Control Engineering 2143, 217230.

Wang, L., Chu, J., Mao, W., 2009. A condition-based replacement and spare
provisioning policy for deteriorating systems with uncertain deterioration to
failure. European Journal of Operational Research 194, 184205.
Yahyaei, M., Banisi, S., Hadizadeh, M., 2009. Modication of SAG mill liner shape
based on 3-D liner wear prole measurements. International Journal of Mineral
Processing 91, 111115.
Yu, G., Qiu, H., Djurdjanovic, D., Jay, L., 2006. Feature signature prediction of a boring
process using neural network modeling with condence bounds. International
Journal of Advanced Manufacturing Technology 30, 614621.
Zhang, S., Ganesan, R., 1997. Multivariable trend analysis using neural networks for
intelligent diagnostics of rotating machinery. Journal of Engineering for Gas
Turbines and Power Transactions of the ASME 119, 378384.