ORIGINAL ARTICLE A new approach to detection of defects in rolling element bearings based on statistical pattern recognition Pavle Stepanic & Ilija V. Latinovic † & Zeljko Djurovic Received: 10 August 2007 / Accepted: 30 January 2009 / Published online: 27 February 2009 # Springer-Verlag London Limited 2009 Abstract The paper presents a new approach to the classification of rolling element bearing faults by imple- menting statistical pattern recognition. Diagnostics of roll- ing element bearing faults actually represents the problem of pattern classification and recognition, where the key step is feature extraction from the vibration signal. Character- ization of each recorded vibration signal is performed by a combination of signal's time-varying statistical parameters and characteristic rolling element bearing fault frequency components obtained through the envelope analysis meth- od. In this way, an 18-dimensional vector of the vibration signal feature is obtained. Dimension reduction of the 18- dimensional feature vectors was performed afterward into two-dimensional vectors representing the training set for the design of parameter classifiers. The classification was performed in two classes, into defective and functional rolling element bearings. Main trait of parameter classifiers is simplicity in their design process, as opposed to classifiers based on neural networks, which employ complex training algorithms. Keywords Feature extraction . Statistical parameters . Fault characteristic frequencies . Envelope analysis method . Dimension reduction . Statistical pattern recognition 1 Introduction Vibrations in a machine occur during rotational motion and during transformation of the rotational motion into transla- tory motion. Two main components of machine rotational motion are rolling element bearings and cogwheels. Vibration signals are frequently used for fault diagnosis of mechanical systems since they carry information on the dynamic state of the mechanical elements themselves. Most modern methods for condition diagnostics of bearings are based on digital processing and analysis of the measured vibration signal from the bearing. All these methods have the same goal—early detection of degrada- tion presence and determining the type of bearing fault, as well as developing a system for monitoring bearing operation. This implies hardware and software development based on state-of-the-art virtual instrument technology and implementation of the Hilbert transform through the technique of envelope analysis [1–3]. The large number of published papers, which discuss this problem, are evidence of its importance. Most of these papers address the same or virtually the same vector of features, which is used as the basis for the classification of healthy and defective bearings. However, the classification approaches differ and, consequently, result in different classification error probabilities. Roughly speaking, two main approaches can be identified among these methods. The first is based on the modeling of vibrations and/or their amplitudes, where classification is made using parametric classification methods [4, 5]. The second group of methods used for fault detection is based on artificial intelligence and soft computing. Artificial neural networks and fuzzy logic are applied to a large extent in solving these types of problems and show that these tools can be used for highly reliable classification. One of the disadvantages of these approaches is that the design process requires the availabil- ity of a very large training set (neural networks) or a large number of parameters, which have to be selected or adjusted (fuzzy systems) to obtain good classification results [6, 7]. A somewhat different approach to this Int J Adv Manuf Technol (2009) 45:91–100 DOI 10.1007/s00170-009-1953-7 † Ilija V. Latinovic – deceased P. Stepanic : I. V. Latinovic : Z. Djurovic (*) Faculty of Electrical Engineering, University of Belgrade, Belgrade, Serbia e-mail: zdjurovic@etf.bg.ac.yu