ng

Maintenance schedulingMill liners

efuld priatis te mipsuct

lation between the input and output variables. The performance of the neural network model is very con-

y usingwns anrvalsnsure sent feechani

Ore grinding mills are heavy duty pieces of equipment thatwork 24 h a day in highly abrasive environments. From an eco-nomic point of view, it is important to keep these mills in operationand minimize the downtime for maintenance or repair, because adrop in production caused by a both scheduled and unscheduledstoppages lead to monetary losses. The auto-genus mill which isused in mineral processing is important for particle size reduction

t or measure thebeen consh is optimions for o

replacement and maintenance scheduling to maximize boformance and prots.

A review of approaches to RUL prediction shows that ancial neural network (ANN) is a powerful tool; it can readily addressmodeling problems that are analytically difcult and for whichconventional approaches are not practical, including complexphysical processes with nonlinear, high-order, and time-varyingdynamics and those for which analytic models do not yet exist.Zhang and Ganesan (1997) used a self-organizing neural networkfor multivariable trending of fault development to estimate the Corresponding author. Tel.: +46 920 49 2106.

Minerals Engineering 53 (2013) 18

Contents lists availab

n

elsE-mail address: Farzaneh.Ahmadzadeh@ltu.se (F. Ahmadzadeh).pact on planning maintenance activities, spare parts provision,operational performance, and protability (Jardine et al., 2006; Al-tay and Green, 2006; Elwany and Gebraeel, 2008; Wang et al.,2009; Kim and Kuo, 2009; Papakostas et al., 2010).

dotiya and Lundberg, 2012), the ability to predicliners wear without stopping the mill has notTherefore, the main focus of the present researcof wear measurement, as well as RUL predict0892-6875/$ - see front matter 2013 Elsevier Ltd. All rights reserved.http://dx.doi.org/10.1016/j.mineng.2013.05.026idered.izationptimalth per-

arti-for researchers; such a system would greatly assist industries, asit is now almost impossible to manually analyze rapidly growingdata to extract valuable decision-making information. In recentdecades, a great deal of research has sought methods of predictingor estimating the remaining useful life (RUL) of critical compo-nents in various industries. Knowing the RUL of an asset has an im-

owner has led to studies of maintenance activities performed onmill liners, such as wear measurement, replacement and mainte-nance scheduling.

Although many studies have been carried out on the effect ofmill liners on grinding mill performance (Cleary, 2001; Santarisiand Almomany, 2005; Kalala et al., 2008; Yahyaei et al., 2009; Dan-1. Introduction

Todays industries are increasingltenance (CBM) to minimize breakdoformance, reduce maintenance inteimprove production efciency, and eof a maintenance systemwith intelligand knowledge accumulation for msistent for data used for training (seen) and testing (unseen). 2013 Elsevier Ltd. All rights reserved.

condition based main-d their impact on per-and consequent costs,afety. The developmentatures in fault detectioncal structures is a goal

and high metal recovery. Among its most critical components arethe liners which protect the mills shell and are used to lift thecharge (ore) inside the mill, thus enhancing grinding performance,Fig. 1. Because their wear inuences the grinding performance inthe context of metal recovery, the mill needs to stop occasionallyfor the maintenance engineer to enter the mill and measure thewear, but each stoppage leads to heavy production losses. The sig-nicant impact of mill liners on the monetary return for the millPrincipal component analysisWear prediction

which inuence wear of the liners. The process parameters were considered as inputs while remainingheight and remaining life of the liners were outputs. The results show remarkably high degree of corre-Remaining useful life prediction of grindineural network

Farzaneh Ahmadzadeh , Jan LundbergLule 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 2013Accepted 30 May 2013Available online 5 August 2013

Keywords:Remaining useful lifeArticial neural network

a b s t r a c t

Knowing the remaining usNow, a mill must be stoppewear, and make the appropthe main aim of this studywithout needing to stop thognize complex relationshinformation-processing str

Minerals E

journal homepage: www.mill liners using an articial

life of grinding mill liners would greatly facilitate maintenance decisions.eriodically so that the maintenance engineer can enter, measure the linerse maintenance decision. As mill stoppage leads to heavy production losses,o develop a method which predicts the remaining useful life of the liners,ill. Because of the proven ability of articial neural networks (ANNs) to rec-between input and output variables, as well as its adaptive and parallelure, an ANN has been designed based on the various process parameters

le at ScienceDirect

gineering

evier .com/locate /mineng

etapiaNota adhesivarestante

etapiaNota adhesivaextremadamente

etapiaNota adhesivaobjetivo

etapiaNota adhesivacaracteristicas

etapiaNota adhesivafallas

etapiaNota adhesivamuy

etapiaNota adhesivaacuerdo

etapiaNota adhesivabuscado

etapiaNota adhesivaenfoques

etapiaNota adhesivafacilmente

operating information, particularly for the wear out phase of linermeasurement. The lack of accurate data for this part of the linercould have hampered the study. However, data requirements werefullled by selecting the shell feed (type1 of the lifter bar (LB1); seeFig. 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).

duction during any mill stoppage is extremely high, it is not eco-

Fig. 1. a Mill liners inside the mill, (b) shell feed lifter bars.

2 F. Ahmadzadeh, J. Lundberg /Minerals Engineering 53 (2013) 18RUL of a bearing system. Shao and Nezu (2000) proposed a newconcept called a progression-based prediction of remaining life toestimate the RUL of a bearing. This concept manipulated the vari-ables determined from online measurements via a compoundmodel of a neural network. Byington et al. (2004) developed a neu-ral network for remaining life predictions for aircraft actuator com-ponents. Yu et al. (2006) presented a neural network model topredict behavior of a boring process during its full life cycle. Maz-har et al. (2007) estimated the remaining life of used componentsin consumer products by using articial neural networks. Runqinget al. (2007) dealt with residual life predictions for ball bearingsbased on a self-organizing map and back propagation neural net-work methods. Similarly, Huang (2007) proposed a method to pre-dict a ball bearings RUL based on a self-organizing map and backpropagation neural network methods. Gebraeel and Lawley(2008) developed a neural network to estimate the RUL of rollingelement bearings by monitoring their vibrations. In this study,ANN has been applied to predict the RUL of the liner with respectto the remaining height and remaining life.

The paper is organized as follows: Section 2 describes the datacollection; 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 basedon life cycle data analysis is hampered due to the unavailability of

nel and engineers from the mining company. These data include

Fig. 2. Generated remaining height and remainithe ore type, ore feed (tonne/h), power (kW), angular speed (% ofcentrifugal critical speed), torque (% of the max torque), wateraddition (m3/h), grinding energy (kW h/tonne), load (tonne). Thisinformation comprised the practical complexities of the grindingprocess inside the mill, including physical explanations of thenomical 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 instal-lation to the next replacement). The solid circles and triangle inFig. 2 show LB1 remaining height and remaining life during one lifecycle; other data are generated by an interpolating technique,Piecewise Cubic Hermite Interpolating Polynomial (PCHIP) andSpline. PCHIP is the best choice for this study, because it nds val-ues 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 mineralRaw process data on the grinding mill, collected from the Bol-

iden mineral databases, were treated to extract the informationused in the models. Information on the important inuencing fac-tors for liner wear was collected in discussions with expert person-Utmost care was taken to ensure that the collected data were asaccurate as possible. There were no data for some periods; in otherperiods, the grinding mill worked only a few hours per day, etc. Wesought to determine why these occurred.

2.1.1. Data from Metso mineralFor this study, an important source of raw data on liner mainte-

nance, inspection and replacement schedule was Metso minerals.The most difcult part was collecting data on liner wear becausethere were fewer measurements. Because total value of lost pro-ng life data by PCHIP interpolating method.

etapiaNota adhesivaaeronave

etapiaNota adhesivaobstaculizado

etapiaNota adhesivafalta

etapiaNota adhesivaprecisin

etapiaResaltado

energy consumption, power and speed uctuation, etc. during dif-ferent wear phases of the mill liners.

3. Prediction of remaining useful life by articial neuralnetwork

As every component operates under different conditions, theactual operating life is dependent on the real conditions of useand, therefore, the operating conditions must be taken into accountwhen determining the life of the product and/or its components. Inother words, the remaining useful life is a function of the actual life(LA) under given operating conditions of use. But real conditions orprocess data will affect the wear out process. The unique nature ofeach data type, process and wear out data, requires specializedanalysis techniques. In this study, ANN has been employed to ndnonlinear or linear relationships between input (process data) andoutput (CM) variables.

3.1. Data preparation

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 andlinks produce a certain ANN model, suitable for certain tasks.

A Multi-Layer Perceptron (MLP) is a kind of feed-forward ANNmodel consisting of three adjacent layers; the input, hidden andoutput layers. Each layer has several PEs. MLPs learn from inputoutput samples to become clever i.e. capable of giving outputsbased on inputs which it has not seen before. The learning processemploys a learning algorithm, during which the MLP develops amapping function between the inputs and outputs. Basically, in alearning process, the input PEs receive data from the external envi-ronment and pass them to the hidden PEs, which are responsiblefor simple, yet, useful mathematical computations involving theweight of the links and the input values. The results from the hid-den PEs are mapped onto appropriate threshold function of each PEand the nal outputs are produced. The output values then becomeinputs to all PEs in the adjacent layer (either the second hiddenlayer or the output layer), and the computation processes are re-peated throughout the layers until nally, output values are pro-

F. Ahmadzadeh, J. Lundberg /Minerals Engineering 53 (2013) 18 3As mentioned earlier, two important factors showing remaininglife, namely, height (mm) and volume (life (%)) of the shell feed(LB1) were considered output variables (see Fig. 1b the remainingvolume 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 in-put variables.

Two life cycles of the lifter bar were combined, the rst from2008-04-15 to 2009-09-23and the second from 2009-12-22 to2011-04-10. We gathered 886 sets of data in both cycles. At rstcycle 80% (347 sets) of the data were used for training and 20%(87 sets) for testing the network. The testing data was grouped inthe 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 infor-mation-processing structures with the ability to build functionalrelationships between data and to provide powerful tools for non-linear, multidimensional interpolations. This aspect of neural net-works makes it possible to capture and interpret the existinghighly complex nonlinear relationships between input and outputFig. 3. Architecture of the pduced at the output PEs. At this stage, an output error value iscalculated by computing the difference between the MLPs andthe actual outputs.

The entire training process is iterative in nature, and stopswhen an acceptably small error is achieved. At completion of alearning process, the MLP should be able to give output solution(s) for any given set of input data based on the generalized map-ping that it has developed Junita and Brian (2008).

The architecture of the proposed three layer network consists oftwo hidden layers of sigmoid transfer function followed by an out-put layer of a linear transfer function. Hidden layers with nonlineartransfer functions allow the network to learn nonlinear and linearrelationships between input and output variables. The linear trans-fer function in the output layer lets the network produce outputsoutside the range [1, 1].

The number of inputs to the proposed network is given by thenumber of available inputs or process data (ore type, ore feed,power, speed, torque, water addition, grinding energy, load); thenumber of neurons in the output layer is constrained to two, asthe output required two parameters, remaining height and remain-ing life, of the liner. The number and size of layers between net-work inputs and the output layer are determined by testingseveral combinations of numbers of layers and various numbersroposed MLP Network.

of neurons in each layer. Each of the selected combinations istested with several different initial conditions to guarantee thatthe proposed model is the best solution. The resulting networkconsists of nine inputs, two hidden layers of 25 and 50 neuronsrespectively, and two nodes at output layers. The architecture ofthe proposed neural network is shown in Fig. 3.

3.3. Principal component analysis

Most of the real-world data samples used to train articial neu-ral networks (ANNs) consist of correlated information from over-lapping input instances. Correlation in sampled data normallycreates confusion during the learning process and, thus, degradesthe generalization capability. The performance of a MLP depends

(the training set) of proper network behavior. The training set con-sists of inputs and the corresponding correct outputs (targets). Oneof the most powerful learning algorithms, the LevenbergMarqu-ardt algorithm (Haykin, 1999), has been used to train networks.In function approximation problems, this algorithm is consideredto have the fastest convergence. Although the network remembersthe training examples, however, it does not learn to generalize tonew situations.

The MATLAB function trainbr, used to train the proposed net-work, has a built-in procedure, Bayesian regularization, designedto overcome over-tting. This technique has been documented asa better generalization procedure for function approximation prob-lems. For the most efcient training, input data are pre-processedbefore training. The selected training function trainbr works bestwhen the network inputs and targets are scaled to fall approxi-

4 F. Ahmadzadeh, J. Lundberg /Minerals Engineering 53 (2013) 18on its generalization capability which, in turn, is dependent onthe data representation. A set of data presented to a MLP oughtnot to consist of correlated information; correlated data reducethe distinctiveness of data representation, introducing confusionto the MLP model during the learning process and producing amodel with low generalization capability to resolve unseen data.This suggests a need to eliminate correlation in the sample data be-fore they are presented to a MLP. This can be achieved by applyingthe principal component analysis (PCA) technique.

PCA involves a mathematical procedure that transforms a num-ber of (possibly) correlated variables into a (smaller) number ofuncorrelated variables called principal components. Basically, Juni-ta and Brian (2008) explained that the PCA technique consists ofnding linear transformations V=(v1, v2, v3, . . .,yp) of the originalcomponents X=(x1, x2, x3,. . .,xQ) that have a property of beinguncorrelated. In other words, the V components are chosen in sucha way that v1 has maximum variance; v2 has maximum variancesubject to being uncorrelated with v1, and so forth. The rst stepin the PCA algorithm is to normalize the components so that theyhave zero mean and unity variance. Then, an orthogonal method isused to compute the principal components of the normalized com-ponents see also, Hong and Wu (2012) and Campbell and Atchley(1981).

In this research, PCA was used to eliminate those componentscontributing less than 5% to the total variation in the data set.The results after applying the PCA indicated signicant redundancyin the training data set, as the principal component analysis re-duced 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 over-tting or over-training. The training style is the supervised learn-ing in which the rules are supplemented with a set of examplesFig. 4. Applying the PCA and reduction of thmately 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 testedfor its learning and generalization capabilities. The test for itslearning ability was conducted by testing its ability to produce out-puts for the sets of inputs (seen data) used in the training phase.The networks outputs had a correlation coefcient of about0.99987 with the desired (actual) outputs. The test for the net-works generalization ability was carried out by investigating itsability to respond to the input sets (unseen data) not included inthe training process.

The results of the proposed neural network showed a remark-ably high degree of correlation between the input and output vari-ables. Clearly, the neural network model could handle the complexnonlinear interrelationships between variables. Furthermore, theperformance of the neural network model was very consistentfor both training and test data. Finally, there was not a consider-able difference in the networks output when trained with eitherseen or unseen data; the proposed model accurately approximatedthe inputoutput function.

4. Results and discussion

Remaining life estimation using nonlinear inputs is far morecomplicated than using linear inputs, especially in the case of anintricate mixture of uctuating and unpredictable trends. Note thatthe results produced by the proposed neural network for remain-ing life assessment are associated with higher levels of certaintybecause statistical analysis techniques were employed.

The best available functions and procedures were utilized topre-process the inputs, train the network and post-process theoutputs of the proposed model. The blue line (variation/noise) inFig. 5 shows the prediction of the ANN in remaining height ande inputs vector dimension from 9 to 5.

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 5

iner6 F. Ahmadzadeh, J. Lundberg /Mremaining life of the liners while the red line (steady variation)shows real outputs. This shows the high ability of the network toproduce outputs for the 80% sets of inputs (347 data sets in the rstcycle) used in the training phase. The gure shows a high correla-tion between actual and predicted output. For generalization abil-ity, 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 in-cluded in the training process; see Fig. 6. A high correlation be-tween the networks predicted output and the actual output isclearly visible. Note that the x-axis in Fig. 7 are showing the dateswhen the measurements are performed for both cycles. Stops ofthe mills are included and thus the x-axis are not showing onlythe total running time of the mills. It can be discussed if the totalrunning time of the mill should be presented in the x-axis insteadof actual date when the measurements are performed. However,from the maintenance point of view it is more practical to use thedate based presentation, since otherwise it is difcult to keep trackof the actual wear at specic time. But in Figs. 5 and 6 because of thenature of selecting the training and testing data (80%:20%) the x-axis are just the data points. It is why (for example) in 87th datapoint the height has come down to around 200 mm in Fig. 6 whileat the same data point in Fig. 5 it is around 350 mm. And it is be-cause the input parameters in the training phase are not the sameas the inputs for test phase. Thus it can be concluded that the neuralnetwork is capable of predicting of the remaining life of the liners.

Another test of the proposed networks ability to predictremaining height and remaining life was training the designedANN with whole data sets belonging to the rst life cycle (434 datasets) and testing its predicting ability for all data in the second cy-cle. The results in Fig. 7 show a remarkably high correlation be-tween real and predicted outputs.

To measure the proposed networks performance, we calculatedthe relative error for its learning and generalization capability. As

Fig. 7. Prediction of ANN for remaining heigals Engineering 53 (2013) 18shown in Figs. 8ac, the maximum relative error was less than6% and 4% for remaining height for seen and unseen data respec-tively. As shown in Fig. 8bd, it was less than 10% for the remaininglife for both seen and unseen data. The network predicted theremaining height and remaining life of the liners with accuracygreater than 90%. Therefore, the proposed model can approximatethe inputoutput function with high accuracy.

5. Conclusions

This study presents a comprehensive approach called ANNwhich can provide remaining life estimates with higher levels ofcertainty. Degradation and condition monitoring data were ana-lyzed by developing a three layer feed-forward back-propagationneural network, consisting of nine inputs which reduced to veafter applying PCA, two hidden layers of 25 and 50 neurons respec-tively, and two nodes at output layers, remaining height (mm) andremaining life (%). The results had 90% accuracy, and the perfor-mance of the neural network model was very consistent for bothtraining data and test data. This is a critical advance in the mainte-nance procedure for grinding mill liners; it allows better under-standing of the mills service requirements and can predict theremaining life of the liners while they remain in operation; it isnot necessary to stop the mill for every maintenance activity, pre-venting high monetary losses.

In summary, the main advantage of the proposed neural net-work is, this methodology does not require disassembly of the lin-ers or stoppage of grinding mills to make decisions on replacementor maintenance. In addition, it does not require expensive andsophisticated equipment for data recording and analysis. More-over, the ANN model can accommodate a wide range of input vari-ables with complex and nonlinear input trends/patterns. However,

ht and remaining life for two life cycles.

inerF. Ahmadzadeh, J. Lundberg /Mit requires large amounts of data for training and tuning. Conse-quently, the dynamic nature of the proposed methodology opensavenues for future study of life cycle data in different categoriesof products.

Acknowledgements

The authors would like to thank Boliden, Metso Mineral for sup-porting this research and permission to publish this article. Specialappreciation is extended to the operating, maintenance engineersJan Burstedt (Boliden), Lars Furtenbach and Magnus Eriksson (Met-so) for sharing their valuable experience, knowledge, and data toimprove 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 networkmethodology 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 chargecomposition. International Journal of Mineral Processing. 63, 79114.

Fig. 8. Relative error for predicted remaining height and remaininals Engineering 53 (2013) 18 7Dandotiya, R., Lundberg, J., 2012. Economic model for maintenance decision: a casestudy for mill liners. Journal of Quality in Maintenance Engineering 18, 7997.

Elwany, A.H., Gebraeel, N.Z., 2008. Sensor-driven prognostic models for equipmentreplacement and spare parts inventory. IIE Transactions 40, 629639.

Eskandari, H., Rezaee, M.R., Mohammadnia, M., 2004. Application of multipleregression and articial neural network techniques to predict shear wavevelocity from wireline log data for a carbonate reservoir, South Iran. CanadianSociety of Exploration Geophysicists CSEG Recorder 29, 4248.

Gebraeel, N.Z., Lawley, M.A., 2008. A neural network degradation model forcomputing and updating residual life distributions. IEEE Transactions onAutomation 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 hybridprincipal component analysis network. Energies 5, 47114725.

Huang, R., 2007. Residual life predictions for ball bearings based on self-organizingmap and back propagation neural network methods. Mechanical Systems andSignal Processing 21 (1), 193207.

Jardine, A.K.S., Lin, D., Banjevic, D., 2006. A review on machinery diagnostics andprognostics implementing condition-based maintenance. Mechanical Systemsand Signal Processing 20, 14831510.

Junita, M., Brian, S., 2008. Improved neural network performance using principalcomponent analysis on Matlab. International Journal of the Computer theInternet and Management 16, 18.

Kalala, J.T., Breetzke, M., Moys, M.H., 2008. Study of the inuence of liner wear onthe load behavior of an industrial dry tumbling mill using the discrete elementmethod (DEM). International Journal of Mineral Processing 86, 3339.

Kim, K.O., Kuo, W., 2009. Optimal burn-in for maximizing reliability of repairablenon-series systems. European Journal of Operational Research 193, 140151.

g life for training and testing phase (seen and unseen data).

Mazhar, M.I., Kara, S., Kaebernick, H., 2007. Remaining life estimation of usedcomponents in consumer products: life cycle data analysis by Weibull andarticial 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 SupportSystems 48, 604612.

Runqing, H., Lifeng, X., Xinglin, L., Richard, L., Hai, Q., Jay, L., 2007. Residual lifepredictions for ball bearings based on self-organizing map and backpropagation neural network methods. Mechanical Systems and SignalProcessing 211, 193207.

Santarisi, N.S., Almomany, R.M., 2005. Mathematical modeling of wear rate of non-repairable parts and their replacement strategies: cement mill liners as a casestudy. 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 ofSystems and Control Engineering 2143, 217230.

Wang, L., Chu, J., Mao, W., 2009. A condition-based replacement and spareprovisioning policy for deteriorating systems with uncertain deterioration tofailure. European Journal of Operational Research 194, 184205.

Yahyaei, M., Banisi, S., Hadizadeh, M., 2009. Modication of SAG mill liner shapebased on 3-D liner wear prole measurements. International Journal of MineralProcessing 91, 111115.

Yu, G., Qiu, H., Djurdjanovic, D., Jay, L., 2006. Feature signature prediction of a boringprocess using neural network modeling with condence bounds. InternationalJournal of Advanced Manufacturing Technology 30, 614621.

Zhang, S., Ganesan, R., 1997. Multivariable trend analysis using neural networks forintelligent diagnostics of rotating machinery. Journal of Engineering for GasTurbines and Power Transactions of the ASME 119, 378384.

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

Remaining useful life prediction of grinding mill liners using an artificial neural network1 Introduction2 Data collection2.1 Data sources2.1.1 Data from Metso mineral2.1.2 Data from Boliden mineral

3 Prediction of remaining useful life by artificial neural network3.1 Data preparation3.2 ANN architecture3.3 Principal component analysis3.4 Training and performance testing

4 Results and discussion5 ConclusionsAcknowledgementsReferences