Fault diagnosis of ball bearings using continuous wavelet transform

Abstract

Bearing failure is one of the foremost causes of breakdown in rotating machines, resulting in costly systems downtime. This paper presents a methodology for rolling element bearings fault diagnosis using continuous wavelet transform (CWT). The fault diagnosis method consists of three steps, firstly the six different base wavelets are considered in which three are from real valued and other three from complex valued. Out of these six wavelets, the base wavelet is selected based on wavelet selection criterion to extract statistical features from wavelet coefficients of raw vibration signals. Two wavelet selection criteria Maximum Energy to Shannon Entropy ratio and Maximum Relative Wavelet Energy are used and compared to select an appropriate wavelet for feature extraction. Finally, the bearing faults are classified using these statistical features as input to machine learning techniques. Three machine learning techniques are used for faults classifications, out of which two are supervised machine learning techniques, i.e. support vector machine (SVM), artificial neural network (ANN) and other one is an unsupervised machine learning technique, i.e. self-organizing maps (SOM). The methodology presented in the paper is applied to the rolling element bearings fault diagnosis. The Meyer wavelet is selected based on Maximum Energy to Shannon Entropy ratio and the Complex Morlet wavelet is selected using Maximum Relative Wavelet Energy criterion. The test result showed that the SVM identified the fault categories of rolling element bearing more accurately for both Meyer wavelet and Complex Morlet wavelet and has a better diagnosis performance as compared to the ANN and SOM. Features selected using Meyer wavelet gives higher faults classification efficiency with SVM classifier.

Introduction

Rolling element bearings are used in a wide variety of rotating machineries from small hand-held devices to heavy duty industrial systems and are primary cause of breakdowns in these machines. There are several vibration and acoustic measurement methods have been used for the detection of defects in rolling element bearings [1]. However, the complex and non-stationary vibration signals with a large amount of noise make the challenging in fault detection of rolling element bearings, especially at the early stage. Therefore, development of effective maintenance strategies and novel diagnosis procedures are needed using different signal analyzing procedures for features extraction and soft computing techniques in order to avoid the system shutdowns, and even catastrophes involving human fatalities and material damage.

To analyze vibration signals, different techniques such as time, frequency and time–frequency domain are extensively used. Samantha and Balushi [2] have presented a procedure for fault diagnosis of rolling element bearings through artificial neural network (ANN). The characteristic features of time-domain vibration signals of the rotating machinery with normal and defective bearings have used as inputs to the ANN. Lei et al. [3] have proposed a method for intelligent fault diagnosis of rotating machinery based on wavelet packet transform (WPT), empirical mode decomposition (EMD), dimensionless parameters, a distance evaluation technique and radial basis function (RBF) network. The effectiveness of wavelet-based features for fault diagnosis of gears using support vector machines (SVM) and proximal support vector machines (PSVM) has been revealed by Saravanan et al. [4]. Yang et al. [5] have proposed a method of fault feature extraction for roller bearings based on intrinsic mode function (IMF) envelope spectrum. Li et al. [6] have shown that the feature vectors obtained by the FFT, wavelet transform, bi-spectrum, etc., can be used as fault features and the HMMs as the classifiers to recognize the faults of the speed-up and speed-down process in rotating machinery. Fault diagnosis of turbo-pump rotor based on support vector machines with parameter optimization by artificial immunization algorithm has been done by Yuan and Chu [7]. Various artificial intelligence techniques are used with wavelet transforms for fault detection in rotating machines [8], [9], [10], [11], [12], [13], [14].

An extensive comparative study concerning the performance of SVM against 16 other popular classifiers, using 21 different data sets, is carried out by Meyer et al. [15]. For classification, simple statistical procedures and ensemble methods proved very competitive, mostly producing good results without the inconvenience of delicate and computationally expensive hyper parameter tuning. For regression tasks, neural networks, projection pursuit regression and random forests often yielded better results than SVMs. The results verify that SVM classifiers rank at the very top among these classifiers, although there are cases for which other classifiers gave lower error rates [15]. Kankar et al. [16] have conducted a comparative experimental study for the effectiveness of ANN and SVM in fault diagnosis of ball bearings and concluded that the classification accuracy for SVM is better than of ANN. Based on the comparison and recommendation of previous studies, authors have employed SVM and ANN for bearing faults classification. In many situations, it is not easy to collect training data set because of routine maintenance and periodically repairs. To solve this problem, authors have also used SOM because unlike SVM and ANN, SOM-based approach has the practical advantage of learning and producing fault classifications without any supervision.

Previous research articles have highlighted the advantages of wavelet transforms when applied to fault diagnosis. In present work, a methodology is proposed for selection of most appropriate wavelet and to determine scale corresponding to characteristic defect frequency based on wavelet selection criterion. These raw signals are divided into 2⁷ sub-signals, i.e. 128 scales in seventh level of decomposition to convert the complex vibration signals into simplified signals with more resolution. Six different wavelets are considered with each 2⁷ sub-signals, i.e. 128 scales. Two wavelet selection criteria Maximum Energy to Shannon Entropy ratio and Maximum Relative Wavelet Energy are used and compared to select an appropriate wavelet for feature extraction. Statistical features are calculated from continuous wavelet coefficients and are fed as input to machine learning techniques, i.e. support vector machine (SVM), artificial neural network (ANN) and self-organizing maps (SOM). The results showed that the proposed methodology can extract useful features from the original data and dimension of original data can also be reduced by removing irrelevant features.