13th World Congress in Mechanism and Machine Science, Guanajuato, M exico, 19-25 June, 2011

A17_603

Vibration-based diagnostics of rolling element bearings: state of the art and challenges
C. Nataraj Department of Mechanical Engineering Villanova University Villanova, PA 19085, USA Abstract Fault identication and estimation is an important and necessary step in Condition Based Maintenance. In general, there are two prevalent methods that are used for this purpose, data driven techniques and model based techniques. Data driven techniques use data collected from experiments, learn about the system and then use this knowledge to infer the systems current state of health; these tend to be more in the applied arena and are favored for implementation in the industry. Model based techniques use models to derive the knowledge that can be used to determine the machines condition; these tend to be more fundamental and are pursued often as a research topic to gain insights into the physics underlying the observed phenomena. Each of these methods has its own strengths and limitations. This paper reviews the state of the art in this area focusing on defects in rolling element bearings, and identies the challenges to be tackled. Finally, a new paradigm is proposed using an illustrative example.
Keywords: rotating machinery, review, diagnostics, condition based maintenance, rolling element bearings, defects

Karthik Kappaganthu Advanced Engineering Cummins Engine, Inc. Columbus, IN 47201, USA

Fig. 1. Diagnostics methodology

I. Introduction With increasing emphasis on efcient, safe, reliable and protable operation of machinery there is a strong need for accurate diagnostics and prognostics. Diagnostics is an essential task in Condition Based Maintenance which increases the longevity of the system and keeps the usually expensive maintenance costs to a minimum. Rotating components are among the most critical components in many machines. The ability to condently determine the state of these systems and predict failures is essential for condition based maintenance. Rotating machines have multiple sources of failure chief among which is bearing failure. Bearings are the load carrying components and are a prominent source of nonlinearity. In this paper, we give an overview of diagnostic methods and outline some new concepts with a special emphasis on detection and estimation of faults in rotor-bearing systems on rolling element bearings. Machinery diagnostics is an important research area that requires inter-disciplinary expertise. Typically machines operate in complicated, uncertain and varying environ nataraj@villanova.edu

ments. A good diagnostic methodology is expected to work accurately in these tough conditions. This necessitates intelligent use of techniques from diverse domains such as machinery dynamics, signal processing, statistical analysis and machine intelligence. Owing to the very complex nature of the problem, the current solutions leave much to be desired. Hence there is a continual need for new paradigms for diagnostics and prognostics. Figure 1 shows the basic steps involved in any diagnostic methodology. Sensors are used to record various signals from the machine but these signals cannot be directly used as such signals are often very large in size and noisy. Signal processing techniques are hence used to derive useful and compact information about the system from these measurements. These information packets are called features and the process is called feature extraction. The features extracted are used to make inferences about the system. Machines with defects would be expected to behave differently from machines without defects and features from a machine with defects will presumably follow a pattern different from the features from a machine without a defect. Computational intelligence algorithms are often used to understand this pattern and infer the state of the system. From the changes in these patterns and features, the future state of the machine is estimated, usually solved with a parameter estimation algorithm.
1

13th World Congress in Mechanism and Machine Science, Guanajuato, M exico, 19-25 June, 2011

A17_603

Fig. 2. Tasks of diagnostics and prognostics

Fig. 3. Motivation schematic

A. Tasks of Diagnostics and Prognostics Typically the tasks of a diagnostic and prognostic methodology are as follows and are shown in Figure 2. 1. Detect any change in performance. 2. Determine if the change is a degradation in the system. 3. Determine the cause of degradation. 4. Identify the amount of degradation. 5. Estimate the remaining useful life. These tasks need to be performed in a sequence. Initially, when there is no reason to expect failure (e.g., just after a scheduled maintenance), it is sufcient to determine if the performance is nominal. This is typically performed using anomaly detection algorithms. At this stage the deviation from a nominal performance is measured. Deviations from nominal system behavior could be due to some temporary change in the operating conditions or due to wear, which is expected over time but not regarded as degradation. The anomaly detection algorithm needs to take these into consideration before agging the anomalies as faults. When an anomaly is determined as a fault, the fault identication algorithm is activated. A machine typically consists of many components each of which could have multiple failure modes. A fault identication algorithm needs to correctly identify the faulty components and the failure modes. Based on the failure mode and the faulty component an appropriate fault monitoring algorithm needs to be applied. When there is sufcient condence of knowledge of the failure and type of failure, prognostics is initiated. There are two tasks to be performed in this phase: rst, to identify and track the amount of fault and second, to understand the progression and make an estimate of the Remaining Useful Life (RUL) of the machine taking into consideration the future usage of the machine. Rotating machines are composed of different subsystems interacting with each other in a nonlinear fashion; changes in either of these components can affect the overall performance signicantly. Due to their inherent nonlinear behavior, analysis of these machines is complicated and it is difcult to derive simple rules indicating and predicting failures.
2

Fig. 4. Inner race defect in rolling element bearings

Referring to Fig. 3, two machines belonging to a same class have certain similarities; these similarities are captured in the physics-based models. Further, each machine based on its environment has a certain individuality; this individuality is captured in the data. Also models generalize the system so that efcient algorithms based on models can be applied to various systems with minor modications. Thus, using both models and data could increase the efciency and accuracy in performing diagnostics and prognostics. This suggests that smart integration of models and data would better predict the performance of the machine. Rotating machines have many components with bearings being one of the key load carrying components. Rolling element bearings are among the most commonly used bearings in high speed applications. They are one of the prominent sources of nonlinearity and often the cause of failure. Bearings themselves have multiple failure modes and have interesting nonlinear behavior. In this paper, we we will use detection of the bearing inner race fault as an illustrative example; more examples can be found in [35]. Figure 4 shows an example of an inner race defect. Details of the model are not provided here because of lack of space. II. Modeling Modeling and analysis of rotating machines has been one of the important research topics in machinery dynamics for quite a long time. This is because, apart from being complicated with multiple sources of nonlinearity, these systems are extensively used in many industries, and often in harsh conditions. Rotating systems have many sources of excitations and nonlinearity and they usually have a residual unbalance which is an external time dependent force on the system. One of the primary sources of nonlinearity in rotating systems is the restoring force from the bearing. In

13th World Congress in Mechanism and Machine Science, Guanajuato, M exico, 19-25 June, 2011

A17_603

[33], [32], [84] the sub and super harmonic excitation of rotating systems is studied. A general approach for nding periodic solutions of rotating systems on nonlinear supports has been developed in [54], [55]. Rolling element bearings are an important class of bearings which are used in many high speed applications. They consist of many moving components namely the inner race, outer race, cage and rolling elements. A detailed introduction to rolling element bearings is available in [27], [49], [51]. Rolling element bearings are highly nonlinear and have three important causes of nonlinearity. They are, the nonlinear restoring forces between various curved surfaces in contact, radial clearance between the races, and defects. The restoring forces between curved surfaces are modeled using Hertzian contact theory. In most of the current models, clearance is modeled as a dead band that is subtracted from the effective displacement of the shaft. Using these as basis [73], [26], [28], [44], [66] have analyzed rotating shafts on rolling element bearings and reported interesting behavior. Many previous studies predicted nonlinear phenomena because of clearance. [83], [11], [71] studied vibrations of shafts with radial clearance. Nonlinear phenomena like bifurcation and chaos have been reported in rolling element bearings [72], [52], [22]. Sub-harmonics, intermittency and quasi-periodic routes to chaos have been reported in these bearing models. The effect of clearance and nonlinear phenomena have been studied using various numerical techniques [20], [40], [21] and nonlinear phenomena like quasiperiodicity [11], [7] and chaos [79], [41], [17], [37] have been reported. Models of localized defects in rolling element bearings have included them as pits in the race or as impulse forces. In [56] the effect of defects on the rotor bearing system has been modeled as the change in the deformation of the rolling element bearing when it enters a pit and more recently in [66] as impulses exerted on the shaft. In [73], [44] a two degree of freedom model of a shaft on a rolling element bearing is considered and the defects are modeled as smooth pits whose shape is an assumed curve. In this model the deformation in the rolling element is a function of the position of the rolling element in the pit. In [18] the deformation in rolling element is derived from the equilibrium position of the rolling element bearing and is used to obtain the forces on the shaft. Clearly, models can have rich and broad information that could and should be mined for bearing diagnostics. Outlined here are some possible ways of integrating this useful information using prevalent statistical and computational intelligence techniques. A. Feature Extraction for Diagnostics and Prognostics Feature extraction for bearing failure, using many signal processing techniques, has been studied extensively and numerous methods have been developed that are sensitive to
3

particular defects. Among the rst feature extraction techniques for rolling element bearing fault detection were the time domain techniques. Rolling element bearings with faults tend to show higher peak to peak vibration compared to a healthy bearing [78], [6], and this has been used as a diagnostic technique. Frequency domain methods are among the most used feature extraction techniques for bearing fault detection. When the rolling element enters a defect, an impulse acts on the casing which excites the structural resonances. The impulse is exerted at a frequency with which the rolling elements enter the defect, which can be calculated from the geometry of the bearing and rotating speed [27], [56], [28]. Frequency domain techniques use these excitations to detect the defects in the bearing. The frequency component associated with a race defect is called the corresponding race ball pass frequency. The rotation of the cage also produces some frequency components. The other frequency components present in a typical bearing signal are the 1X response usually due to rotating unbalance, and its harmonics and subharmonics. The presence of harmonics and sub-harmonics indicates nonlinear behavior in general. Note that a system with a healthy bearing is also nonlinear in nature. The Fast Fourier Transform (FFT) is the most common method to extract the frequency components in a signal. The spectrum usually contains a peak at the defect frequency. However, this is not always clearly observable because of slip and masking by other stronger vibrations. This is more of a problem when one wishes to detect inner race defects because the excitation signals need to travel through the rolling element, and the casing, and then be detected by the sensor by which time the signal is masked by other excitations. In order to overcome this problem, an envelope spectrum technique is often used. The impulses excitations are amplitude modulated and can be recognized as side bands in the frequency spectrum and can be seen as peaks in the envelope spectrum [50], [67]. In order to nd the envelope spectrum the signal is band pass ltered around a frequency that has the maximum signal to noise ratio, and then the Hilbert transform is used to nd the envelope spectrum. There are many techniques to nd the central frequency and range for the band pass lter [25], [74], [80]. Spectral kurtosis is often used to select this band because of the simplicity and ease of use [4]. The Discrete Wavelet Transform (DWT) is a method for obtaining the time-frequency information of the signal. These methods are useful to extract the transients in the signal and are hence popular for inner race defect detection. More information about the wavelet transforms can be found in [10], [58], [59]. Some of the recent work on bearing diagnostics using DWT are [62], [9], [63], [14], [82], [19], [53], [60]. Another set of novel feature extraction techniques is based on nonlinear signal processing and complexity mea-

13th World Congress in Mechanism and Machine Science, Guanajuato, M exico, 19-25 June, 2011

A17_603

sures such as Approximate Entropy and Lyapunov Exponents. Morphology and Approximate Entropy based techniques are being used increasingly in bio-medical signals and for fault detection purposes in machines. In the medical eld such techniques have been efciently used to analyze EEG and CTG signals [13], [1], [76]. Entropy measures for estimating the similarity between signals have been used in [90] for biomedical applications and in [45] for foreign exchange time series analysis. In the analysis of machinery, measures based on these techniques have been used to diagnose gearboxes [34] and rolling bearing defects [46], [47], [86]. [61] used similar methods to extract envelopes for impulsive-type periodic systems. [85] performed a detailed study on approximate entropy for detecting degradation in signals and demonstrated it to detect the severity of defects in rolling element bearings. Although these feature extractions are novel, they are mostly data driven and are highly dependent on the operating conditions, because of which it is difcult to generalize these techniques. To overcome this problem cross-sample entropy which can be evaluated relative to a model has been recently proposed [38]. Cross-sample entropy is a type of nonlinear feature extraction technique that can be used to measure the similarity between the measurements and the simulations. Cross-sample entropy is an extension of approximate entropy [65] which is an estimate of Kolmogorov entropy. These techniques are complexity measures that capture the information creation in a time series [57]. Cross-sample entropy captures the match between parts of two signals; the greater the similarity in the two signals, the lower the cross-sample entropy. Another advantage of cross-sample entropy over approximate entropy is that it has a smaller bias and is more consistent. More detailed information can be found in [70]. As an example, we present some results from a recent study [35] in which a model based method was developed to identify the severity of outer race defects. For this purpose, two bearings of different fault magnitudes were used and their corresponding models developed. Cross-sample entropy was used to compare the similarity between the simulations and measurements. These comparisons showed that if the simulation and the measurement were similar, the cross-sample entropy was lower and vice versa. Further, the variation of cross-sample entropy with the variation in operating conditions was minimal. This is because the changes in the operating conditions are directly updated in the parameters that are used in the simulations. Because of this direct relationship there is no need to recalibrate the algorithm and study the effects when operating conditions change. This is a very useful property, as, in practice, the data driven methods need to be calibrated when the operating conditions change which is usually a time consuming and difcult process. The method developed provides a simple and accurate method to identify severity of faults. In a practical setting, models at various levels of severity can
4

1.2 1.1

Defect Free

Speed

1 0.9 0.8

1.2 1.1 1 0.9

Outer Race Defect

2.5 2

Covariance

Cross Sample Entropy

Fig. 5. Clustering: defect-free and inner race defect

be simulated and stored as a database. Subsequent measurements can be compared to the simulations in the database using the method developed in this research to evaluate the severity of the defect in the system. The database of models and simulations can be easily maintained and upgraded based on the system and diagnostic accuracy necessary. Another interesting idea is to use clustering to match some feature from the model and the defective data. The key task in this approach is to develop a good measure for the degree of match or closeness. This measure must be able to identify subtle differences in the signal because of varying defect magnitudes and types. At the same time, it should be robust to noise and uncertainties. If an accurate enough model is used then the data from a defect-free system would be closer to the model estimate than the data from a system with a defect. Such clustering of data is used for effective classication of data. An example is shown in Fig. 5 from [35]. In this example, the simulation of the model at various operating conditions and defect parameters was used to develop novel model based methods to detect defects in bearings. The measurements from bearings with and without defects were compared to this simulation using a variety of features obtained using time domain and morphological signal processing algorithms. The comparison was performed using a support vector machine. The clustering together of data from similar systems was used for classication. The algorithms performance was excellent for the data considered. The results showed the feasibility and advantages of using model based features. The few errors that were observed occurred at lower speeds where the accuracy of the model parameters is lower. B. Feature Selection Feature selection is an important part of a diagnostic algorithm. The performance of a diagnostic methodology greatly depends on the quality of the features being used. Typically, features are used as a set in a diagnostic algo-

13th World Congress in Mechanism and Machine Science, Guanajuato, M exico, 19-25 June, 2011

A17_603

rithm; in this case, care should be taken that a right combination of features are used and that they contain as much pertinent information about the fault as possible. When the operating conditions in which the machine operates vary, the features should contain the information about these variations in operating conditions, implicity or explicitly. Although there are many studies on feature selection, there are only a few that analyze bearing defects quantitatively. Sugumaran et al. [77] use decision trees to select features that provide good performance using a proximal Support Vector Machine using time domain features like skewness, kurtosis etc. In [48], Principal Component Analysis (PCA) was used to develop a set of features that improved the performance of both supervised and unsupervised learning machines. In this study, features were generated using time domain features like skewness, kurtosis etc., and using frequency domain techniques like amplitudes at fault frequencies, and time-frequency domain techniques like wavelet transforms. In both these studies the outputs of the algorithms were feature sets that provided the best performance for the given data set but there was no ordering and comparison of the features. Some other general computational intelligence based algorithms that can be used for feature selection are evolutionary algorithms [24], [69]. Recently, mutual information, an information theoretic approach has been proposed to perform feature selection in rolling element bearings [36]. This method is also useful to integrate model based features and data driven features. Mutual information based feature selection has been used in computer security, face recognition and biomedicine [31], [87]. C. Classiers Bearing defect detection is usually formulated as a classication problem. Classiers such as Articial Neural Network (ANN) [2], [42], Adaptive Neuro-Fuzzy Interference System (ANFIS) [5], and Support Vector Machines (SVM) [91], [43] have been used for bearing defect detection. Articial neural networks are information processing systems based on biological processing systems. Neural networks consist of many computing units called perceptrons which are arranged in layers and are interconnected as shown in Fig. 6; this is the reason ANNs are sometimes referred as Multi Layered Perceptrons (MLP). Each neuron is characterized by input, output, weight and an activation function. The output of each neuron is the processed output of the activation function using the weighted sum of the input as the argument. ANNs are supervised learning machines [29] which can be trained using representative data. Training is performed by back propagation during which the weights are adjusted to minimize the error between the outputs and the ANN predictions. This trained network can then be used to predict the output of any data with properties similar to the training
5

Fig. 6. Structure of Articial Neural Networks

data. Support vector machines are statistical learning theory based supervised learning machines. These are feedforward linear learning machines and were pioneered by Vapnik [81]. Support vector machines are approximate implementations of the principle of structural risk minimization. These machines provide good generalization performance on pattern classication [29]. More detailed analysis of support vector machines can be found in [29], [75]. III. The problem with feature selection In the literature, several data driven features have been researched extensively. Data driven feature extraction techniques can be categorized into time, frequency and timefrequency domain techniques. Each of these techniques has its own sensitivity to various types of defects and operating conditions. The choice of the features also depends on the system in which the bearings are present, and the conditions in which the bearing is expected to perform. Often, one feature is insufcient and multiple features need to be used for efcient classication. Thus, the interaction between these features is important for good diagnostic performance. Also, too few or too many features would degrade the classiers performance. Given some sample data, this feature set selected by the method needs to have the highest information content as well as the right set of features for maximum accuracy. Although there are many studies on feature selection there are only a few that analyze bearing defects quantitatively. In this research, an information theoretic approach is used to determine the optimal set of features [64]. This technique quanties the quality of features as the mutual information content between features and the state of the bearing (faulty or healthy). Mutual information is a statistical measure that correlates different random variables [16]. It can be calculated from the probability distribution between the random variables [12]. Mutual information can be used to compare the features with each other and rank them accordingly. The advantage of using this information theoretic approach is that it is independent of the classier used. Also,

13th World Congress in Mechanism and Machine Science, Guanajuato, M exico, 19-25 June, 2011

A17_603

among the features, some of these might have similar information among themselves. Hence, using such features together increases the redundancy and uncertainty, and therefore degrades the performance. The information theoretic approach addresses this important issue of interaction of features with each other for classication purposes and provides a set that performs better cumulatively. It also provides an ordered set of features that can be used to rank features and increase the efciency of classication. Thus, using mutual information addresses three important issues. First, it illustrates a quantitative statistical method to compare features for detection of two kinds of faults in bearings irrespective of the classier; second, it provides an ordered optimal feature set comprised of features from different domains that together provide good classication accuracy; and third, it provides guidelines about the features that need to be considered for fault monitoring purposes. In a recent study [39], time (skewness, kurtosis) [78], frequency (FFT, envelope) [6], [67], [88] and time-frequency (discrete wavelet transform) [9], [53], [88] domain features are compared with each other. It should be noted that the list of features considered here is not exhaustive. We have chosen these because these are the most basic techniques that are fundamentally different from each other. There are clearly many other techniques which are variations and hybrids of these techniques. There are also many other novel signal processing techniques based on nonlinear signal processing techniques [61], [23], demodulation [30], [8], [3], [68] and empirical mode decomposition [15], [89], [43]. These are not a part of this study; however, the proposed methodology could certainly be used to analyze and compare these techniques. Mutual information was used to rank and compare features for detecting outer and inner race defects in rolling element bearings. Some of the basic features for bearing fault diagnostics from time, frequency and time-frequency domain features were extracted and ranked. This is a procedure which could easily be used to study the relative performance of any other advanced or newly developed features for bearing fault detection. The utility of this method has been demonstrated by applying it on a data set generated from experiments. Using mutual information the features compared were ranked and ordered. The ordering of the features provided many insights into the system and the operating conditions. The FFT magnitude at bpf o turned out to be the most important feature for outer race defect classication and DWT detail energy at level four was ranked highest for inner race defect classication. This indicates the effect of slip and other interfering disturbances in the case when the outer race defects were minimal. Also, the procedure effectively showed that these features alone were not capable of good classication performance; information from other frequency and time-frequency domain features was needed. Another in6

teresting observation was the presence of FFT and envelope magnitudes at /2 and 2. These features are a measure of the nonlinearity in the system, which means that the bearings with defects excite different nonlinear behavior of the system compared to the non-defective systems as is known from mathematical modeling. This suggests that specic information about the defect is to be corroborated with other information about the system captured in speed, FFT magnitudes etc., to obtain good classication performance. Some of the inferior features such as DWT approximate skewness, kurtosis, etc., were ranked lower in the ordered feature set. In the case of the inner race defect, the envelope magnitude at ball pass frequency ranked higher than the FFT magnitude at the same frequency. This validates the ability of the procedure to reject inferior features automatically. Clearly, such ability is invaluable in cases when the relative importance of the various features is not easily apparent. Another advantage of ordering the features is that it has provided a suggestion of features that are better suited for dealing with fault monitoring and prognostics for the system. When an ANN was used to determine the optimal feature subset from the ranked feature set, the ANN classication accuracy showed trends similar to the mutual information content. This illustrates that the mutual information based feature ranking and feature selection could be a useful tool that could be used to select the features that are best suited for the system and conditions in which the bearing is operating. It greatly reduces the analysis required for developing a good bearing diagnostic system. Using a validation data set, an optimal feature subset was extracted. The optimality condition was the mean minimum classication error using ten ANN simulations. The process of selecting the optimal feature set was easy because the features were ordered to provide the best possible performance as a set corroborating each others information content and reducing redundancy. The optimal feature set in this case had twelve features. The method showed excellent performance on a test data set as shown in Figure 7. The other advantage of this method is that this feature set guarantees high performance over a range of speeds. A nal word of caution is that, since this is a data driven technique, the performance of the features could be significantly different for data that is signicantly different from our data; of course, this is an inherent limitation with all data driven techniques. Real world systems tend to be very different and much more complicated than laboratory systems! With that caveat in mind, it is still possible to state with a reasonable degree of certainty that this study has established the potential benet of using mutual information to make objective decisions for designing critical bearing diagnostic systems.

13th World Congress in Mechanism and Machine Science, Guanajuato, M exico, 19-25 June, 2011

A17_603

Start
105 100 95 90

Models Operating Conditions

Data Collection

Feature Extraction Data Driven Model Based

Tune Model Parameters

Accuracy

Validation Set

Training Set Rank Features Based on Mutual Information

Test Set

85 80 75 70 65 60

Inner Race Defect Defect Free Mean

i=1

Ordered Feature Set i=i+1 No is Yes i=Length of Features Set Evaluate Performance Select Feature Subset With Highest Accuracy Stop

ANN
0 5 10 15 20 25 30 35

Accuracy

Feature Order
Fig. 7. Classication performance on validation data for inner race defect classication
100

Fig. 8. Algorithm for feature integration

IV. A new paradigm We propose a new paradigm where the model based features and some of the common data driven features for fault identication in bearings are integrated to obtain the optimal combination of features by careful feature ranking and selection. The mutual information based method for feature ranking followed by a classier for feature selection can be used to generate an optimal set of features that provides optimal diagnostic performance. The optimal feature set is designed not only to defects but also to identify fault severity. The aim of this study is to understand quantitatively the interaction of the data driven features and model based features, and their effect on classication performance. Further, we use this study to identify the advantages of model based features for bearing fault detection. The owchart of the process is provided in Fig. 8 . In the data collection step vibration data is collected from a system with a faulty bearing and a defect-free bearing over a span of rotating speeds and load, and is used for training, validation and testing of the algorithm. The faulty bearing has either a localized small or large outer race defect. From the data, various data driven and model based features are extracted. The operating condition parameters of the model, namely, speed and load are determined based on the collected data and the corresponding models are used to extract the model based features. This is followed by a greedy search algorithm which is used to rank the features based on the mutual information. Next, the validation set is used to extract an optimal feature subset for classication using ANN as the classier. The feature subset selection is performed incrementally using the ordered feature set obtained in the previous stage.
7

90

80

Accuracy (%)

70 Data Driven Features Model Based Featues Combined Features

60

50

40

30

10

15

20

25

30

Features

Fig. 9. Classication performance

The subset with the best ANN classication performance is the optimal solution that integrates model based features and data driven features. This optimal feature subset is then used to test the performance using the test set data. From Figure 9 it can be seen that the data driven features outperform the model based features when considered independently. Also, no single data driven or model based feature performs well by itself. In both cases, the performance increases steeply after the rst couple of features. It is interesting to note the decline in the performance of the model based features after the rst three features. This is probably because the rest of the model based features are redundant. From all the possible feature sets from the three ordered feature sets, the best possible performance is obtained when both data driven features and model based features are used. This performance is achieved when the rst ten features from the ordered feature set are used. Further the perfor-

13th World Congress in Mechanism and Machine Science, Guanajuato, M exico, 19-25 June, 2011

A17_603

mance is signicantly improved when model based features are used. The best accuracy using data driven features is 93.4% which improves to 99.3% when the optimal feature set suggested by the algorithm is used. V. Closing Thoughts This paper presented a fairly comprehensive review of model and data based diagnostic algorithms for rotating machinery with a special emphasis on defects in rolling element bearings. In addition, a new paradigm is outlined as a potential future research area implementing integration of the information in models and data efciently. This novel idea of integration of information has been developed and validated to detect, and to identify type and severity of localized defects in rolling element bearings. Rolling element bearings have been chosen for this study as they are the load carrying elements in many high speed machinery, one of the primary sources of nonlinearity and often the cause of failure. The approach proposed in this paper requires the use of many algorithms from different domains including nonlinear dynamic modeling, pattern recognition, machine intelligence and signal processing, and is quite challenging. A method has been developed to integrate the information in these features with information in data driven features. For this fusion, a mutual information based feature ranking and feature selection technique is used. First, this technique was used to rank and select some of the basic data driven features for bearing fault detection. These features were used in this study as they are among the commonly used features for bearing diagnostics and are representative of many other techniques which are some combinations or improvements of these basic techniques. The various data driven features extracted were ranked based on their mutual information content with various types of defects i.e., fault identication. Next, this ordered feature set was used to select the subset of features that provided the best classication performance. The ordering of the features provided many insights into the physics of the system with its operating conditions. One important observation here was that features such as FFT magnitude at harmonics of rotating speeds that characterize the nonlinear behavior of the system were ranked among the top. This suggests the importance of nonlinearity for accurate diagnostics and prognostics. Using this method, the information in various data driven features was efciently fused to obtain the right set of features that provide optimal diagnostic performance. Next, using this information theoretic method, an optimal feature set that has the information from both the models and the data is generated to efciently and accurately identify the severity of defects in rolling element bearings at varying operating conditions. Using the information theoretic method, the optimal set of features were selected so that the features contain relevant information about the defect and are not redundant with each other. This ensures
8

the efcient use of information available in both models and data to obtain superior diagnostic performance. This is evident from the results that show that for the same system, the diagnostic performance is signicantly better when both data driven and model based features are used. Further, the optimal combination of these features can be efciently chosen by rst ranking the features using mutual information and then performing selection, which greatly reduces the number of iterations necessary to obtain the optimal feature set. Many future directions suggest themselves. Better mathematical models would for example consider the exibility of the shaft and analyze the effect of defects on such systems. As a future work it would be interesting to determine the resolution of the model based severity estimation algorithm developed by many authors and to improve it using better statistical clustering techniques. For this purpose, accuracy and variance of the model parameters estimation and their relation to severity estimation needs to be studied. Also, it would be interesting to use other signal comparison techniques or similarity measures that are typically used in image recognition to estimate the closeness of a measurement to a simulation and to determine the parameters. Another approach to determine the fault parameters using the model simulations could be based on the fundamentals used in model predictive control and estimation that model the uncertainty along with the dynamics and optimally estimate uncertainty parameters. Novel approaches using particle lters can also be developed to use physics based models to perform fault identication, monitoring and prognostics. Currently these approaches are being used in systems where simplistic models are available. It needs to be studied if they can be used efciently with high delity models. References
[1] [2] R.U. Acharya, O. Faust, N. Kannathal, T. Chua, and S. Laminarayan. Non-linear analysis of eeg signals at various sleep stages. Computer Methods and Programs in Biomedicine, 80:3745, 2005. I. E. Alguindigue, A. Loskiewicz-Buczak, and R. E. Uhrig. Monitoring and diagnosis of rolling element bearings using articial neural networks. IEEE Transactions on Industrial Electronics, 40:209217, 1993. J. Altmann and J. Mathew. Multiple band-pass autoregressive demodulation for rolling-element bearing fault diagnosis. Mechanical Systems and Signal Processing,, 15:963977, 2001. J. Antoni and R. B. Randall. The spectral kurtosis: Application to the vibratory surveillance and diagnostics of rotating machines. Mechanical Systems and Signal Processing, 20:308331, 2006. Makarand S. Ballal, Zafar J. Khan, Hiralal M. Suryawanshi, and Ram L. Sonolikar. Adaptive neural fuzzy inference system for the detection of inter-turn insulation and bearing wear faults in induction motor. IEEE Transactions on Industrial Electronics, 54:250258, 2007. A. Barkov, N. Barkova, and J. Mitchell. Condition assessment and life prediction of rolling element bearings. Sound and Vibration, 28:1017, 1995. R. F. Beatty and M. J. Hine. Improved rotor response of the uprated high pressure oxygen turbopump for the space shuttle main engine. Journal of Vibrations, Acoustics, Stress and Reliability in Design, 111:163169, 1989.

[3] [4] [5]

[6] [7]

13th World Congress in Mechanism and Machine Science, Guanajuato, M exico, 19-25 June, 2011

A17_603

[8] [9]

[10] [11] [12] [13] [14] [15]

[16] [17] [18] [19] [20] [21]

[22] [23]

[24]

[25] [26]

[27] [28] [29] [30] [31] [32]

I. S. Bozchalooi and Ming Liang. Parameter-free bearing fault detection based on maximum likelihood estimation and differentiation. Measurement Science & Technology, 20:065102, 2009. Iain S. Cade, Patrick S. Keogh, and M. Necip Sahinkaya. Fault identication in rotor/ magnetic bearing systems using discrete time wavelet coefcients. IEEE/ ASME Transctions on Mechatronics, 10(6):648657, December 2005 2005. Y. T. Chan. Wavelet Basics. Kluwer Academic Publihsers, 1995. D. W. Childs. Fractional frequency rotor motion due to clearance effects. Transactions of the ASME Journal of Engineering Power, 104:533536, 1982. T. M. Cover and J. A. Thomas. Elements of Information Theory. John Wiley and Sons, New York, 1991. L. Diambra, J.C. Bastos, and C.P. Malta. Epileptic activity recognition in eeg recording. Physica A, 273:495505, 1999. Abderrazek Djebala, Nouredine Ouelaa, and Nacer Hamzaoui. Detection of rolling bearing defects using discrete wavelet analysis. Meccanica, 43(3):339348, 2008. Hongbo Dong, Keyu Qi, Xuefeng Chen, Yanyang Zi, Zehengiia He, and Bing Li. Sifting process of emd and its application in rolling element bearing fault diagnosis. Journal of Mechanical Science and Technology, 23:2000, 2009. Richard O. Duda, Peter E. Hart, and David G. Stork. Pattern Classication. Wiley-Interscience, New York, 2001. F. F. Ehrich. Observation of subcritical superharmonic and chaotic response in rotordynamics. Journal of Vibration and Acoustics, 114:93100, 1992. I. K. Epps. An Investigation into Vibrations Excited by Discreet Faults in Rolling Element Bearings. PhD thesis, University of Canterbury, 1991. Yanhui Feng and Fernando S. Schlindwein. Normalized wavelet packets quantiers for condition monitoring. Mechanical Systems and Signal Processing, 23:712723, 2009. G. T. Flowers and F. Shengwu. A study of the inuence of bearing clearance on lateral coupled shaft/disk rotordynamics. Journal Of Engineering For Gasturbine, 115:279286, 1993. L. Frosini and P. Pennacchi. Detection and modelling of rotor eccentricity in electrical machines - an overview. In Eighth International Conference on Vibrations in Rotating Machinery - IMechE Conference Transactions, 2004. S. Fukata, E. H. Gad, T. Kondou, T. Ayabe, and H. Tamura. On the radial vibrations of ball bearings (computer simulation). Bulletin of the JSME, 28:899904, 1985. Phanindra Garimella and Bin Yao. Robust model-based fault detection using adaptive robust observers. In Proceedings of the 44th IEEE Conference on Deccision and Control and the European Control Conference, pages 30733078, Serville, Spain, December 2005. H. Guo, L. B. Jack, and A. K. Nandi. Automatic feature extraction for bearing fault detection using genetic programming. Eighth International Conference on Vibrations in Rotating Machinery - IMechE Conference Transactions, 2:363372, 2004. Lie Guo, Jin Chen, and Xinglin Li. Rolling bearing fault classication based on envelope spectrum and support vector machine. Journal of Vibration and Control, 15:13491363, 2009. T. C. Gupta, K. Gupta, and D. K. Sehgal. Nonlinear vibration analysis of an unbalanced exible rotor supported by ball bearings with radial internal clearance. ASME Conference Proceedings, 2008(43154):12891298, 2008. T. A. Harris. Rolling Bearing Analysis. Wiley-Interscience, 4 edition, 2002. S. P. Harsha, K. Sandeep, and R. Prakash. Non-linear dynamic behaviors of rolling element bearings due to surface waviness. Journal of Sound and Vibration, 272(3-5):557 580, 2004. Simon Haykin. Neural Networks: A comprehensive foundation. Prentice Hall, 1999. Xiaoyi Hu, Qingfu He, and Huasheng Wang. Rolling bearing fault detection based on svd denoising and stft demodulation method. China Railway Science, 29:95100, 2008. Jinjie Huang, Yunze Cai, and Xiaoming Xu. A hybrid genetic algorithm for feature selection wrapper based on mutual information. Pattern Recognition Letters, 28:18251844, 2007. Y. Ishida, T. Ikeda, and T. Yamamoto. Nonlinear forced oscillations caused by quartic nonlinearity in a rotating shaft system. Journal of Vibration and Acoustics, 112:288297, 1990.

[33] Yukio Ishida and Toshio Yamamoto. Forced oscillations of a rotating shaft with nonlinear spring characteristics and internal damping (1/2 order subharmonic oscillations and entrainment). Nonlinear Dynamics, 4:413431, 1993. [34] J.D. Jiang and J. Chen. The application of correlation dimension in gearbox condition monitoring. Journal of Sound and Vibration, 224:529541, 1999. [35] Karthik Kappaganthu. Machinery diagnostics using an integrated approach with models and data. PhD thesis, Villanova University, 2010. [36] Karthik Kappaganthu and C. Nataraj. Feature selection for fault detection in rolling element bearings using mutual information. Journal of Vibration and Acoustics, under review, 2010. [37] Karthik Kappaganthu and C. Nataraj. Nonlinear modeling and analysis of a rolling element bearing with a clearance. Communications in Nonlinear Science and Numerical Simulation, In Press, Corrected Proof:, 2011. [38] Karthik Kappaganthu, C. Nataraj, and B. Samanta. Model based bearing fault detection using sample entropy features and support vector machines. Journal of Sound and Vibration, 2009 (under review). [39] Karthik Kappaganthu, C. Nataraj, and Biswanath Samanta. Optimal feature selection for detection of defects in rolling element bearings. Mechanical Systems and Signal Processing, 2009 (under review). [40] Y. B. Kim and S. T. Noah. Bifurcation analysis for a modied Jeffcott rotor with bearing clearances. Nonlinear Dynamics, 1:221241, 1990. [41] Y. B. Kim and S. T. Noah. Quasiperiodic response and stability analysis for a nonlinear Jeffcott rotor. Journal of Sound and Vibration, 190:239253, 1996. [42] B. Li, M. Y. Chow, Y. Tipsuwan, and J. C. Hung. Neural-networkbased motor rolling bearing fault diagnosis. IEEE Transactions on Industrial Electronics, 47:10601069, 2000. [43] Meng Li and Ping Zhao. The application of wavelet packet and SVM in rolling bearing fault diagnosis. In IEEE International Conference on Mechatronics and Automation, 2008. [44] A. Liew, N. Feng, and E. J. Hahn. Transient rotordynamic modeling of rolling element bearing systems. Journal of Engineering for Gas Turbines and Power, 124(4):984991, 2002. [45] Li-Zhi Liu, Xi-Yuan Qian, and Heng-Yao Lu. Cross-sample entropy of foreign exchange time series. Physica A: Statistical Mechanics and its Applications, 389(21):4785 4792, 2010. [46] D. Logan and J. Mathew. Using the correlation dimension for vibration fault diagnosis of rolling element bearings-i. basic concepts. Mechanical Systems and Signal Processing, 10:241250, 1996. [47] D. Logan and J. Mathew. Using the correlation dimension for vibration fault diagnosis of rolling element bearings-ii. selection of experimental parameters. Mechanical Systems and Signal Processing, 10:251264, 1996. [48] Arnaz Malhi and Robert X. Gao. PCA-based feature selection scheme for machine defect classication. IEEE Transactions on Instrumentation and Measurement, 53:15171525, 2004. [49] P.D. McFadden and J.D. Smith. Model for the vibration produced by a single point defect in a rolling element bearing. Journal of Sound and Vibration, 96:6982, 1984. [50] P.D. McFadden and J.D. Smith. Vibration monitoring of rolling element bearings by the high frequency resonance techniquea review. Tribology International, 117:310, 1984. [51] P.D. McFadden and J.D. Smith. The vibration produced by multiple point defects in a rolling element bearing. Journal of Sound and Vibration, 98:263273, 1985. [52] B. Mevel and J. L. Guyader. Routes to chaos in ball bearings. Journal of Sound and Vibration, 162:471487, 1993. [53] K. Mori, N. Kasashmi, T. Yoshioka, and Y. Ueno. Prediction of spalling on ball bearings by applying discrete wavelet transform to vibration signals. Wear, 8:195162, 1996. [54] C. Nataraj. Periodic solutions in rotor dynamic systems with nonlinear supports: a general approach. PhD thesis, Arizona State University, 1987. [55] C. Nataraj and H. D. Nelson. Periodic solutions in rotor dynamic systems with nonlinear supports: a general approach. American Society of Mechanical Engineers Journal of Vibration, Acoustics, Stress, and Reliability in Design, 111:187193, 1989.

13th World Congress in Mechanism and Machine Science, Guanajuato, M exico, 19-25 June, 2011

A17_603

[56] C. Nataraj and Robert Gerard Pietrusko. Dynamic response of rigid rotors supported on rolling element bearings with an outer raceway defect. ASME Conference Proceedings, 2005(47381):12491261, 2005. [57] Ali H. Nayfeh and Balakumar Balachandran. Applied Nonlinear Dynamics. Wiley-VCH, 1995. [58] D.E. Newland. Wavelet analysis of vibration, part 1: theory. Transactions of the American Society of Mechanical Engineers, Journal of Vibration and Acoustics, 116:409416, 1994. [59] D.E. Newland. Wavelet analysis of vibration, part 2: wavelet maps. Transactions of the American Society of Mechanical Engineers, Journal of Vibration and Acoustics, 116:417425, 1994. [60] D.E. Newland. Ridge and phase identication in the frequency analysis of transient signals by harmonic wavelets. Transactions of the American Society of Mechanical Engineers, Journal of Vibration and Acoustics, 121:149155, 1999. [61] N. G. Nikolaou and I. A. Antoniadis. Application of morphological operators as envelope extractors for impulsive-type periodic signals. Mechanical Systems and Signal Processing, 17:11471162, 2007. [62] Hasan Ocak, Kenneth A. Loparo, and Fred M. Discenzo. Online tracking of bearing wear using wavelet packet decomposition and probabilistic modeling. Journal of Sound and Vibration, 302:951 961, 2007. [63] Yuna Pan, Jin Chen, and Lei Guo. Robust bearing performance degradation assessment method based on improved wavelet packet -support vector data descriptions. Mechanical Systems and Signal Processing, 23:669681, 2009. [64] Hanchuan Peng, Fuhui Long, and Chris Ding. Feature selection based on mutual information : Criteria of max-dependency, maxrelevance and min-redundancy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27:12261238, 2005. [65] S.M. Pincus. Approximate entropy as a complexity measure. Chaos, 5:110117, 1995. [66] Ahmad Rafsanjani, Saeed Abbasion, Anoushiravan Farshidianfar, and Hamid Moeenfard. Nonlinear dynamic modeling of surface defects in rolling element bearing systems. Journal of Sound and Vibration, 319(3-5):1150 1174, 2009. [67] R. B. Randall and Y. Gao. Extraction of modal parameters from the response of power cepstrum. Journal of Sound and Vibration, 176:179193, 1994. [68] R.B. Randall and N. Sawalhi. Signal processing tools for tracking the size of a spall in a rolling element bearing. In IUTAM Symposium on Emerging Trends in Rotor Dynamics, 2009. [69] Michael L. Raymer, William F. Punch, Erik D. Goodman, Leslie A. Kuhn, and Anil K. Jain. Dimensionality reduction using genetic algorithms. IEEE Transactions on Evolutionary Computation, 4:164 171, 2000. [70] Joshua S. Richman and J. Randall Moorman. Physiological timeseries analysis using approximate entropy and sample entropy. American Journal of Physiolgical Heart and Circulatory Physiology, 278:H2039H2049, 2000. [71] S. Saito. Calculation of nonlinear unbalance response of horizontal jeffcott rotors supported by ball bearings with radial clearances. Transactions of ASME, Journal of Vibration, Acoustics, Stress, and Reliability In Design, 107:416420, 1985. [72] A. Sankaravelu, S. T. Noah, and C. P. Burger. Bifurcation and chaos in ball bearings. Nonlinear and Stochastic Dynamics, 78:313325, 1994. [73] N. Sawalhi and R. B. Randall. Simulating gear and bearing interactions in the presence of faults part 1. the combined gear bearing model and simulation of localized bearing faults. Mechanical Systems and Signal Processing, 22:19241951, 2008. [74] N. Sawalhi and R. B. Randall. Semi-automated bearing diagnostics - three case studies. In Comadem Conference, Faro, Portugal, Faro, Portugal, June, 2007. [75] John Shawe-Taylor and Nello Christianini. Kernel Methods for Pattern Analysis. Cambridge University Press, 2004. [76] M.G. Signorini, G. Magenes, S. Cerutti, and D. Arduini. Linear and nonlinear parameters for the analysis of fetal heart rate signal from cardiotocographic recordings,. IEEE Transactions on Biomedical Engineering, 50:365374, 2003. [77] V. Sugumaran, V. Muralidharan, and K. I. Ramachandran. Feature selection using decision tree and classication through proximal sup-

[78] [79] [80]

[81] [82] [83] [84] [85] [86] [87]

[88] [89] [90]

[91]

port vector machine for fault diagnostics of roller bearing. Mechanical Systems and Signal Processing, 21:930942, 2007. N. Tandon. A comparision of some vibration parameters for condition monitoring of rolling element bearings. Measurement, 12:285 286, 1994. M. Tiwari, K. Gupta, and O. Prakash. Effect of radial internal clearance of a ball bearing on the dynamics of a balanced horizontal rotor. Journal of Sound and Vibration, 238(5):723 756, 2000. P. W. Tse, Y. H. Peng, and R. Yam. Wavelet analysis and envelope detection for rolling element bearing fault diagnosis-their effectiveness and exibilities. Transactions of the ASME. Journal of Vibration and Acoustics, 123:303313, 2001. V. N. Vapnik. Statistical Learning Theory. John Wiley & Sons, 1998. Jian-Da Wu and Chiu-Hong Liu. Investigation of engine fault diagnosis using discrete wavelet transform and neural network. Expert Systems and Applications, 35:12001213, 2008. T. Yamamoto. On the vibration of a shaft supported by bearings having radial clearances. Transactions of the JSME, 21:182192, 1955. T. Yamamoto. On sub-harmonic and summed and differential harmonic oscillations of a rotating shaft. Bulletin of JSME, 4:5158, 1961. R. Yan and R. X. Gao. Approximate entropy as a diagnostic tool for machine health monitoring. Mechanical Systems and Signal processing, 21:824839, 2007. R. Yan and R.X. Gao. Complexity as a measure for machine health evaluation. IEEE Transactions on Instrumentation and Measurement, 53:12371334, 2004. Zhiguo Yana, Zhizhong Wanga, and Hongbo Xieb. The application of mutual information-based feature selection and fuzzy LS-SVMbased classier in motion classication. Computer Methods and Programs in Biomedicine, 90:275284, 2008. A. Ypma. Learning Methods of Machine Vibration Analysis and Health Monitoring. PhD thesis, Delft University, 2001. Yang Yu, Yu Dejie, and Cheng Junsheng. A roller bearing fault diagnosis method based on EMD energy entropy and ANN. Journal of Sound and Vibration, 294:269277, 2006. Tao Zhang, Zhuo Yang, and John H. Coote. Cross-sample entropy statistic as a measure of complexity and regularity of renal sympathetic nerve activity in the rat. Experimental Physiology, 92:659 669, 2007. Zhousuo Zhang, Wenzhi Lv, and Minghui Shen. Active learning of support vector machine for fault diagnosis of bearings. Advances In Neural Networks, 3973:390395, 2006.

10