You are on page 1of 2

PREDICTING RELIABILITY OF HEAVY MINING EQUIPMENT USING

COMPUTATIONAL INTELLIGENCE DYNAMIC MODELS


Gonzalo Acuña (1), Millaray Curilem (2), Beatriz Araya (3), Fernanda Garrido (2) & Francisco Cubillos (1)

(1) Universidad de Santiago de Chile (USACH), Facultad de Ingeniería, Santiago, Chile

(2) Universidad de La Frontera (UFRO), Departamento de Ingeniería Eléctrica,Temuco, Chile

(3) DIRECTIC Company, Santiago, Chile

Summary: In this work the development of nonlinear autoregressive with exogenous inputs models to predict
Mean Time Between Failures (MTBF) and Mean Time to Repair (MTTR) variables is presented. The models were
developed using support vector machine and historical data obtained on a daily basis during 2013 from one heavy
mining equipment of an important copper mine site in Chile. One-step-ahead predictions of the predicted variables
confirmed good performance of the dynamic models. Reliability status of the analysed heavy mining equipment can
be predicted from MTBF and MTTR forecasts thus allowing the maintenance engineer to have not only a static
view of the situation but also a view of its evolution under different scenarios

Keywords: SVM, NARX, Mining Equipment, Reliability, Mean Time between Failures, Mean Time to Repair

1 INTRODUCTION
The fleet of heavy mining equipment corresponds to the most important asset of the mining industry. Maintenance
budgets of these assets are relevant and contribute to a large part of the cost of production. One reason is the poor
precision of maintenance plans which today are based on a simple historical analysis. Having a predictive tool to
forecast availability and other variables may allow improving those maintenance plans. Well known computing
intelligence techniques, like neural networks and support vector machine (SVM), have been extensively used as
classifiers and also as regressive models for developing adequate dynamic models of complex systems. The use
of SVM as a tool for the development of systems to support maintenance has also been reported in the literature
[1,2]. In this work SVM is used as a regressive tool for developing dynamic models for predicting Mean Time
Between Failures (MTBF) and Mean Time to Repair (MTTR) using the Use of Physical Availability (UPA) as an
exogenous input for both dynamic models. Prediction of those variables is a valuable tool for characterizing the
reliability state of the analysed equipment.

2 MATERIAL AND METHODS

2.1 Support Vector Machine

In 1995 Vapnik and his colleagues proposed the classification technique called Support Vector Machine (SVM) [3].
The approach proposed for classification problems, can be adapted to regression. Regression is defined as the
fitting of data into a hypertube of radius ε [4]. The process that finds the optimal hypertube requires the tuning of
some parameters, such as the regularization parameter C, the kernel parameter sigma and the radius of the
hypertube ε. In this work the C and sigma parameters were increased by powers of 2. The exponents took values
from -5 to 15, with step 1. Parameter ε varied between 0 and 0.5 with a step of 0.1. The performance of every
combination was assessed calculating the validation error. The library of tools developed at INSA de Rouen,
France [5] was used to implement the SVR models.

2.2 Development of the dynamic models

Two dynamic models using SVR were developed. Both were built as Nonlinear Autoregressive with Exogenous
inputs (NARX) models with MTBF (or MTTR) as the regressive variable while UPA was used as the exogenous
input. Data correspond to historical information of 3 variables: MTBF, MTTR and UPA obtained in a daily basis
during 2013 for one heavy mining equipment. The used methodology includes splitting the data into 3 sets: training
data set (150 examples); validation data set (100 examples) and test data set (86 examples). The first two sets
were used for adapting SVR parameters while the third served for testing the elaborated models in a one-step-
ahead (OSA) prediction scheme. The order of the models was obtained in a trial and error basis. The final
developed models are presented in Figure 1.

Asset Management Conference 2013 AMBoK ID: ### Page 1


Figure 1 NARX SVR models for MTBF and MTTR variables for the test set

3 RESULTS
Both models were tested using the test data set. OSA predictions for MTBF and MTTR are shown in Figure 2
where yTEST is the experimental data and yOSA the predicted values. The time was normalized between 0 and 1.
nd
The best values for the MTBF model were obtained with two delays (2 order model) while for the MTTR model
three delays (3 order) were necessary. In both cases the tuning process lead to C=211, sigma=25, ε=0.1.
d

Figure 2 Experimental and predicted data for MTBF and MTTR variables for the test set.
The comparison between experimental data and the predictive values, as shown in Figure 2, shows good
performance of the developed models with a SMAPE, Correlation and RMS indexes of 27.96, 0.99 and 0.21
respectively for MTBF and 151.85, 0.96 and 0.16 for MTTR.

4 CONCLUSION – This paper presents two NARX SVM models for predicting variables used for reliability
analysis of heavy mining equipment. The operator can use the predicted MTBF and MTTR variables to
classify in advance an equipment into “critical”, “chronic”, “critical&chronic” and “under control” categories. The
UPA variable showed better effect on regression than others like availability. Both variables required a
d nd
nonlinear model being MTTR more complex to model (3 order) than MTBF (2 order). Although results are
promising this is a preliminary work that should be improved when more data and other relevant variables –
like monitoring conditions, environmental variables, etc …- could be available. The goals are to improve OSA
predictions and also to achieve good multiple step ahead predictions.

5 ACKNOWLEDGEMENTS – The authors acknowledge partial financial support of Chilean Government


under FONDEF-IDeA GRANT CA13I10121 and of the Research Direction of Universidad de La Frontera.

6 REFERENCES
[1] Ding, F., He Z., Zi, Y., Chen, X., Tan, J., Cao, H., Chen, H., 2008, “Application of support vector
machine for equipment reliability forecasting”, Proceedings of the IEEE International Conference on
Industrial Informatics, 526-530.
[2] Zeng, Y., Jiang, W., Zhu, C., Liu, J., Teng, W., Zhang, Y., 2006, “Prediction of equipment maintenance
using optimized support vector machine”, Lecture Notes on Artificial Intelligence, LNAI 4114, 570-579.
[3] Vapnik, V. 1995, The Nature of Statistical Learning Theory, Springer Verlag.
[4] Drucker,H., Burges,C.J.C. , Kaufman, L.,Smola,A. andVapnik, V., 1997, Support vector regression
machines. In M. Mozer, M. Jordan, and T. Petsche, editors, Advances in Neural Information
Processing Systems 9, Cambridge, MA, MIT Press.pp 155-161.
[5] Canu, S., Grandvalet, Y., Guigue, V., Rakotomamonjy, A, 2005, SVM and Kernel Methods Matlab
Toolbox. Perception Systèmes et Information, INSA de Rouen, Rouen, France.

Asset Management Conference 2013 AMBoK ID: ### Page 2

You might also like