Fault Diagnosis of Rolling Element Bearings Using Vibration Signal Analysis – A Soft Computing Approach Dr. Srinivasa Pai P. 1 , Vijay G. S. 2 , Dr. N. S. Sriram 3 1, 3 Dept. of Mechanical Engineering, NMAMIT, Nitte 2 Dept. of Mechanical and Manufacturing Engineering, MIT, Manipal e-mail id: 1 srinivasapai@rediffmail.com; 2 vgs.mpl@gmail.com; 3 nittesriram@yahoo.com Abstract: Vibration signal analysis is a widely used technique for monitoring the condition of rolling element bearings (REB). Artificial Neural Networks (ANN) and Support Vector Machines (SVM) are widely used soft computing techniques in fault diagnosis of REBs, along with Genetic Algorithms (GA). In this paper, ANNs and SVM have been used for fault diagnosis of REBs. Vibration signals collected using accelerometers mounted on the bearing housing, have been used to analyse the bearing condition. Normal bearing and defective bearings (with inner and outer race defect) have been analysed and the vibration signal features namely RMS, 0-P, crest factor, load and speed were used as inputs to train and test two types of ANNs namely Multilayer Perceptron (MLP) and Radial Basis Function (RBF) neural network. The results obtained have been compared with that obtained using SVM. It was found that MLP trained using different training algorithms gave an accuracy of around 85 %, whereas for RBF networks, it was around 60 % and for SVM it was around 73 % on test data. Thus MLP and SVMs were effective, when compared to RBF networks in fault diagnosis of REBs. I. INTRODUCTION REBs are commonly used machine element used in rotating machinery, for transmitting load between two members usually a shaft and housing [1,2]. Defective bearings are a major cause of problems in machines. Most common defects in REBs are the localized defects, which include cracks or pits on the inner race, outer race and rolling elements [2]. These defects when excited during operation of the bearing can generate vibrations. This vibration signal can be monitored to detect the presence of these defects. The vibration signals collected in the vicinity of a bearing assembly contain rich information about the bearing condition. Most of the researchers have used vibration signature analysis techniques for REB fault detection in case of single defect on bearing components [3]. Vibration signal analysis therefore is a powerful diagnostic and troubleshooting tool. Time domain approach is the simplest approach for vibration signal analysis. In time domain analysis, trending of acquired vibration signals with respect to time, monitoring of indices (like mean, range, RMS value, kurtosis and crest factor) helped to identify the presence of defects in REB [4]. ANNs have been widely used for automatic fault detection and diagnosis in machine conditions. Artificial Intelligence (AI) have been increasingly applied to machine diagnosis and have shown improved methods over conventional approaches. Among AI techniques, ANNs are the most widely used for machine diagnosis [5]. Samanta et.al (2003) used ANNs for fault diagnosis of REBs. The features used included RMS, variance, skewness, kurtosis and normalised sixth central moment of the time domain vibration signals. The ANN had two hidden layers and the output layer consists of two binary nodes indicating the status as normal or defective bearings. ANN was found to be very effective [6]. Samanta et.al (2004) have compared the performances of MLP, RBF and probabilistic neural network (PNN) with Genetic Algorithm (GA) based feature selection from time domain vibration signals, for two class (normal or faulty) recognition of REB [7]. Sunil Tyagi (2008) compared ANNs and SVM for fault diagnosis of REBs. Simple statistical features such as standard deviation, skewness, kurtosis etc. of the time domain vibration signals along with peaks of the signal and peak of power spectral density (PSD) were used as input to ANN and SVM. It was found that the performance of SVM was better than ANN and pre-processing of vibration signals using Discrete Wavelet Transform (DWT) enhances the effectiveness of both ANN and SVM classifier [2]. Omar Jose et.al (2006) used wavelet decomposition to extract features from vibration signals and selected features were used as inputs to three types of ANNs trained to identify the bearing conditions at three different rotational speeds. The vibration signals were acquired from a motor-driven experimental system. The developed diagnostic method was able to reliably detect and classify four different bearing fault conditions namely normal bearing, bearing with inner race fault, outer race fault and ball faults [8]. In this paper, ANNs and SVM have been used for fault diagnosis in REBs. Vibration signal (time domain) acquired through a hand held Data Acquisition System (DAQ) and an accelerometer, mounted on the test bearing housing have been