An application of learning machine methods in prediction of wear rate of wear resistant casting parts q Radomir Slavkovic a , Zvonimir Jugovic a , Snezana Dragicevic a,⇑ , Aleksandar Jovicic b , Vladimir Slavkovic c a Technical Faculty Cacak, University of Kragujevac, Svetog Save 65, 32000 Cacak, Serbia b Faculty of Mechanical Engineering, University of Kragujevac, Janjic Sisters 6, 34000 Kragujevac, Serbia c School of Electrical Engineering, University of Belgrade, King Alexander Avenue, 73, 11120 Belgrade, Serbia article info Article history: Received 9 March 2012 Received in revised form 25 December 2012 Accepted 29 December 2012 Available online 11 January 2013 Keywords: Support vector machine Kernel functions Wear rate prediction Optimum chromium content abstract In this paper, a method of floating ball wear rate identification, using two machine-learning techniques Support Vector Machine (SVM) and Improved Support Vector Machine (ISVM) are proposed. Both models are used to relate the wear rate and technological parameters of the wear resistant drip moulding using different kernel functions. The models for determining the wear rate of white iron casting with low chro- mium content (flotation balls), was trained and tested by using the existing exploitation data from the Bor Flotation Plant, Serbia. In order to select the best model parameters the statistical indicators for both models are presented. Results show that the ERBF (SVM) and ERBF+POLY (ISVM) achieved the best clas- sification accuracy compare to other kernels used: the absolute mean error of ERB (SVM) is 5.85%, while the error of ERBF+POLY (ISVM) is 6.67%. The tuned ISVM model with mixture of kernels is able to accu- rately predict the wear rate and can be used to define the optimum chromium content in liquid metal alloys for the casting of flotation balls. Ó 2013 Elsevier Ltd. All rights reserved. 1. Introduction The wear rate of flotation balls depends upon their material, mechanical–chemical characteristics obtained from the casting process and ore composition. The technology of developing flota- tion balls through hardness (HRC) and chemical composition has an influence on the balls’ wear rate. In order to achieve the optimum process of ore milling, it is necessary to establish a relationship between the wear rate and the technological process parameters. For that purpose a machine learning methods have been used i.e. a Support Vector Machine (Gunn, 1998) and improved support vector machine (Jaya, Vinodhini, & Karthik, 2012; Smits & Jordaan, 2002) and then the one that gives the best results was used to manage the casting process of floating balls. SVM is based on statistical learning theory and is a new achievement in the field of data-driven modelling and has been successfully implemented in classification, regression and function estimation (Shi & Gindy, 2007). The concept underlying this algo- rithm is that of observing the relationships that are valid for a finite set of data. By identifying and learning these relationships, SVM ac- quires the characteristic of generalisation, which means that the algorithm will be able to perform predictions for a new data set generated by the same source. Recently, SVM has been widely used to solve various problems in almost all scientific disciplines. Huang, Li, and Gan (2010) analysed SVM as a supervised method for problem classification taking advantage of prior knowledge of tool wear. Cho, Asfour, Onar, and Kaundinya (2005) applied Sup- port Vector Machines for Regression (SVR) to model the power and maximum cutting force in an end milling application. In their investigation, the SVR approach was better than a Multiple Vari- able Regression (MVR) approach. Yang and Shieh (2010) used ma- chine learning approach i.e. SVR for model development that predicts consumers’ affective responses for product form design, through a comparison of two standard kernel functions (polyno- mial and RBF). Li (2009) presents a machine vision and image pro- cessing techniques combining a novel classifier, support vector machine, to detect and classify copper clad laminate surface de- fects, while Shin, Eom, and Kim (2005) used one-class support vec- tor machines in machine fault detection and classification. Based on ground roughness variation during the grind able period, Chiu and Guao (2008) used SVM for classifying the intercepted grinding acoustic emission data. Their SVM model was constructed from the result of a grinding experiment and was found to predict with 85% accuracy. Wang, Jing, and Ren (2007) introduced the application of wavelet packet and SVM method for the classification recognition of the rift wear mark. Malyscheff, Trafalis, and Raman (2002) de- scribe how the support vector algorithm can be modified in order to identify the minimum enclosing zone for straightness and flatness tolerances. To improve the precision of coal reserve 0360-8352/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.cie.2012.12.021 q This manuscript was processed by Area Editor Manoj Tiwari. ⇑ Corresponding author. Tel./fax: +381 32302710. E-mail addresses: vedra.md@open.telekom.rs, snezad@tfc.kg.ac.rs (S. Dragicevic). Computers & Industrial Engineering 64 (2013) 850–857 Contents lists available at SciVerse ScienceDirect Computers & Industrial Engineering journal homepage: www.elsevier.com/locate/caie