Vous êtes sur la page 1sur 11

ARTICLE IN PRESS

Mechanical Systems and Signal Processing 24 (2010) 289299

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing


journal homepage: www.elsevier.com/locate/jnlabr/ymssp

Case History

Automated diagnosis of rolling bearings using MRA and neural networks


n , O. Lara, J.C. Garc a-Prada C. Castejo
MAQLAB Group, Mechanical Dept., Universidad Carlos III, Av. de la Universidad, 30, 28911 Madrid, Spain

a r t i c l e in fo
Article history: Received 29 November 2007 Received in revised form 4 June 2009 Accepted 11 June 2009 Available online 27 June 2009 Keywords: Wavelets Articial networks Fault diagnosis Predictive maintenance Pattern classication

abstract
Any industry needs an efcient predictive plan in order to optimize the management of resources and improve the economy of the plant by reducing unnecessary costs and increasing the level of safety. A great percentage of breakdowns in productive processes are caused by bearings. They begin to deteriorate from early stages of their functional life, also called the incipient level. This manuscript develops an automated diagnosis of rolling bearings based on the analysis and classication of signature vibrations. The novelty of this work is the application of the methodology proposed for data collected from a quasi-real industrial machine, where rolling bearings support the radial and axial loads the bearings are designed for. Multiresolution analysis (MRA) is used in a rst stage in order to extract the most interesting features from signals. Features will be used in a second stage as inputs of a supervised neural network (NN) for classication purposes. Experimental results carried out in a real system show the soundness of the method which detects four bearing conditions (normal, inner race fault, outer race fault and ball fault) in a very incipient stage. & 2009 Elsevier Ltd. All rights reserved.

1. Introduction All rotating machinery used in the modern world can develop faults. Maintenance plans include analyzing the relevant external information of critical components in order to evaluate their internal state. Therefore, bearing failures are the common cause of rotating machinery breakdown. Realtime on-line monitoring can increase early detection and fault diagnosis automation, but a more reliable and faster mathematical method is required. Classication of incipient faults in rolling bearings is an open line of research. There is not much work related to this issue. In the two last years, prestigious journals have published only four or ve papers related to incipient bearing fault diagnosis [1,2,7]. In these papers, data collected from an experimental lab bench do not include radial load, which is the most important force the rolling bearings are designed for. Probably one of the most important elements in industrial machines are shafts, which transmit energy. Radial rolling bearings are designed to support the loads that the shafts transmit: basically, the radial load provided by the masses of different elements connected to the shaft (gears, cranks, etc.) and their own mass, which in industrial machines is not negligible, and a certain degree of axial loads due to misalignments. In this paper, this consideration of loads applied to the rolling bearing has been included in the experimental setup. This difference from other authors makes the defect classication more difcult due to the increment of the contact ellipse between the ball and the race. As can be seen in the developments carried out by Harris [3], the width

Corresponding author. Tel.: +34 916 249 186; fax: +34 916 249 430.

n). E-mail address: castejon@ing.uc3m.es (C. Castejo 0888-3270/$ - see front matter & 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.ymssp.2009.06.004

ARTICLE IN PRESS
290 n et al. / Mechanical Systems and Signal Processing 24 (2010) 289299 C. Castejo

of the contact surface is directly proportional to the square root of the radial load. That is, the higher radial load that the bearing supports, the more ellipse surface contact exists. In these cases, contact between the ball and the race cannot be considered a point, so incipient faults can go in the contact surface, making it difcult to detect them. There are two important stages to implement in the fault diagnosis process: the rst is signal processing, for feature extraction and noise diminishing, and the second one consists of signal classication, based on the characteristics obtained in the previous stage. Most of the research related to bearing fault diagnosis agrees with the use of vibration signature for this purpose, due to the non-stationary characteristics the signals present when a fault occurs in the rolling element bearing operation [1,47]. In recent years, different technologies have been used in order to process signals provided from dynamical systems. Most of the authors classify the analysis of vibration signature in three approaches [5,6]: time domain based on statistical parameters such as mean, root mean-square, variance, kurtosis, etc. [8], frequency domain, where the Fourier transform (FT) [2] and its variations [9] were the most commonly used in the past; and time-frequency analysis such as the wavelet transform (WT) [10]. This last approach is the most commonly used in signatures with non-stationary characteristics. The most classical approaches are the power spectra density (PSD) and demodulation analysis (based on frequency domain). The rst approach gives us an idea of the energy of each frequency peak obtained from fast Fourier transform (FFT). The demodulation approach or envelope analysis consists of obtaining the spectra from the temporal signal envelope (based on the Hilbert Transform). These approaches have been demonstrated to be useful in detecting bearing faults (not in an incipient stage) under laboratory conditions, where all the other sources of faults are reduced or removed. In the case of incipient faults, the amplitude of the spectra is very low and other techniques are needed. The WT has been successfully applied as a fault feature extractor due to the good energy concentration properties. Peng et al. in [11] carried out a bibliographical review of the WT application in the monitoring and fault diagnosis in machines. The main drawback of WT, apart from the selection of the suitable basis function for performing the transformation, is that it is not able to separate the high frequency bands where the information of the machine operating with failure is presented. This problem is solved by using the wavelet packet transform (WPT) proposed by Liu et al. in 1997 [12]. The WPT is a multiresolution analysis (MRA) technique [9] which gives a suitable frequency-band partition. Whereas most of the authors that use WT in failure diagnosis develop a method to discard the less representative coefcients obtained (for example, with the threshold method as Chen et al. develop in [13]) for the next classication step, the wavelet packet coefcients can be directly used as features, and they possess a high sensibility to failures [1]. In summary, many kinds of fault features can be obtained, principally with wavelet coefcients or wavelet energy. Since wavelet coefcients will highlight the changes in signals which often predict the occurrence of the fault, the Wavelet coefcients-based features are suitable for fault detection. However, because slight changes in signals often have little energy, these changes will be easily masked in the wavelet energy-based features. Therefore, the wavelet energy-based features are often not able to detect early faults. Signal processing is a relevant item in a bearing fault diagnosis system. Nevertheless, in order to obtain a monitoring system which concludes the real condition of the rotatory element, a classication system is needed. New trends in fault diagnosis try to develop intelligent classication systems. Preliminary research can be found in [14,15]. Lou et al. in [2] used a fuzzy classier to diagnose faults in bearings, based on the use of the discrete wavelet transform (DWT) as a feature vectors generator. Hu et al. uses an support vector machine (SVMs) ensemble in [1], and the rest of the researchers in the eld use genetic algorithms [16] or neural networks [17,18] as a classication kernel. In the present work, vibration signature from rolling element bearings is processed by means of the WPT and a neural classier in order to detect four bearing conditions. The ow chart of the monitoring procedure proposed in this manuscript is shown in Fig. 1. The methodology proposed performs the diagnosis procedure in a direct way, without developing the detection and identication task developed by other authors [2,1]. In this sense, vibratory signals represent the input to the monitoring system, which are going to be processed at once, in order to obtain relevant information about the component condition or state. Subsequently, a previously trained classier system will provide the diagnosis of the system condition. This methodology reduces the effect of the human factor during the diagnosis process. 2. Experimental setup Vibratory signals have been obtained thanks to the test lab bench presented in Fig. 2. In this bench, developed by the UNED mechanical department, FAG 7206 B single ball bearings were tested. In the gure, starting on the right hand-side, the following elements are visible: axial and radial pneumatic cylinders, the bearing assembly, a B&K 4383 accelerometer with an 8.5 kHz bandwidth, a phototachometer device for RPM measurement, and a transmission pulley directly connected

Fig. 1. Diagnosis procedure ow chart.

ARTICLE IN PRESS
n et al. / Mechanical Systems and Signal Processing 24 (2010) 289299 C. Castejo 291

Fig. 2. Bearing Test bench. UNED Lab.

Fig. 3. Bearing fault: (a) inner race; (b) outer race.

to the motor by a V-belt. Additional acquisition devices are a B&K NEXUS amplier, a DAS-1200 Keithley acquisition card. The sampling rate was set at 5000 Hz, and each acquired signal had 5120 points. Four sets of data were obtained from the experimental system: (i) (ii) (iii) (iv) under normal conditions; inner race faults; outer race faults and ball fault.

A pit 2 mm long was articially induced in the inner or outer race by an electric pen. In the case of the rolling ball, multiple slots in the surface were performed to simulate the acking phenomenon. Bearing faulted races are shown in Fig. 3. The radial and axial loads were 2.5 and 3 bars, respectively. A total of 196 bearings measured were obtained, 49 for each condition at 600 RPM. Parallel studies were performed at 1200 and 1800 RPM. As the literature shows, there are characteristic frequencies for each type of fault [19]. For ball bearings used in the experimental results (FAG 7206B) the frequencies for a running speed of 600 RPM are: BPFI (ball pass frequencyinner): 75.39 Hz. BPFO (ball pass frequencyouter): 54.60 Hz. BSF (ball spin frequencyrolling elements): 23.34 Hz.

ARTICLE IN PRESS
292 n et al. / Mechanical Systems and Signal Processing 24 (2010) 289299 C. Castejo

normal condition 1 0 1 0 0.1 0.2 t (seg) ball fault 15 10 5 0 5 0.3 0.4 15 10 5 0 5

x 105

normal condition

100

200 f [Hz] ball fault

300

400

0.15 0.1 0.05 0 0.05

x 107

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 t (seg) outer race fault

0 x 105

100

200 300 f [Hz] outer race fault

400

500

4 0 1 2 0 0.1 0.2 t (seg) inner race fault 2 0 2 0 0.1 0.2 t (seg) 0.3 0.4 4 2 0 2 0.3 0.4 2 0

0 x 104

100

200 300 f [Hz] inner race fault

400

100

200

300 f [Hz]

400

500

Fig. 4. Experimental data and rst processing: left, raw data from experimental setup for the four bearing conditions (acceleration [m=s2 ]); right, spectra PSD of each signal [mV2 =Hz].

Nevertheless, these frequencies cannot be observed in the spectra during the initial stage of the bearing deterioration progress, that is, when an incipient fault appears. On the other hand, there are other frequencies that appear in the spectra due to the different elements connected. In Fig. 4, the signals obtained from the experimental setup for the four conditions of the rolling bearing and their spectra are presented. In this gure, some frequencies of interest can be seen but, we need a previous preprocessing of the raw data in order to obtain useful information for incipient fault classication, mainly for the last two bearing conditions (outer and inner race faults). The main difference between a laboratory machine and an industrial machine is that the rst one can be considered to be formed by rigid solids, without friction and perfectly tted and balanced, while in the case of real machines, this is not true. As can be seen in Fig. 4, the characteristic bearing condition frequencies have, in general, a very low amplitude in the incipient stage, and they can be masked due to noise and other machine vibrations.

3. Signal processing task In this manuscript, analysis of PSD and envelope-PSD has been performed (see spectra in Fig. 4). The results do not let us predict the incipient fault, due to the number of frequencies presented around the characteristic ones. After evaluating this result, we conclude that, when an industrial machine is operating, there are too many factors to add to the vibration signature, and a previous pass-band lter for eliminating useless information is not easy to implement. A wavelet package lets us perform a rst lter and add the non-stationary characteristic to the signal. Parameters obtained with the level of detail selected let us generate a pattern interesting for classication purposes. The method proposed lets us automate the complete maintenance process in a real industry. Some insights of wavelet analysis for signal processing purposes are now briey presented. The wavelet analysis has the advantage of better performance for non-stationary signals, representing a time signal in terms of a set of wavelets. They are constituted for a family of functions which are derived from a single generating function called the mother wavelet, through dilation and translation processes. Dilation is related to size, and it is also known as a scale parameter while translation is the position variation of the selected wavelet along the time axis.

ARTICLE IN PRESS
n et al. / Mechanical Systems and Signal Processing 24 (2010) 289299 C. Castejo 293

Fig. 5. Subband coding by means of two channels.

The WT cannot be used practically by using analytical equations; therefore, a discretization process is needed. Given f n, the discrete signal to be analyzed, the DWT formulation is given by X f nCj;k n j; kZ2 (1) cj; k
z

where Cj;k n tj=2 Ctj n ks is the mother wavelet. Parameters t and s are dened according to the dyadic scale, as appearing in Eq. (2), where j represents the decomposition level and k the translation factor.

t 2j ;

s k 2j

t40;

s41 j; kZ2

(2)

Multiresolution analysis is the most suitable way to perform the discretization [9]. It consists of recursive lter application, starting by applying a half band low pass lter, which possesses an impulse response hn. The lter ts a mathematical convolution with the signal and the lter impulse response, eliminating all the frequency components over half of the maximum signal frequency. MRA analyzes the signal in different frequency bands, but with different resolutions, having separated the signal in two different data: approximation and detail information. Two functional groups are used to represent the information. They are called scaled functions (f) and wavelet functions (c), which are associated with the low-band and high-band lters, respectively. After ltering, the same number of samples will be achieved, but in half of a frequency band. By applying the Nyquist rule, half of the samples are justied to be eliminated without losing information. This procedure is the rst decomposition level (level 1) and can be expressed mathematically as follows: X xng 2k n (3) yhigh k
n

ylow k

X
n

xnh2k n

(4)

yhigh and ylow are the low (h) and high (g ) band output lter, respectively; the term 2k determines the subsampling by a factor of two, and n is the number of original samples. This procedure, also called subband coding, may be represented by Fig. 5, and it can be repeated for additional decompositions. 3.1. MRA for feature extraction In this work, mother wavelet Daubechies-6 has been selected for processing a number of 49 signals per condition. Fig. 6 shows the comparison of a signal under normal conditions and inner race fault. The approximation level (a5), and ve detail levels (d1d5) are shown for each signal. Based on these plots, the inner race fault and the normal condition show several magnitude differences in several frequency bands, like d3, d4 and d5. Similar results were obtained for ball and outer race faults. The fth level detail coefcients (cD5) have been chosen as characteristic features [20]. Since the articial neural networks (ANN) need as few coefcients as possible for classication tasks, the last decomposition level gives fewer coefcients (18) than others. In addition, this frequency band is visually different in each condition. The feature vectors are post-processed by normalizing in the range 1 1 and they belong to the frequency band (78.125156.25 Hz). In general, for each discrete signal f n we have a characteristic vector Vcf n that can be dened as in Eq. (5): Vcf i n cD5f i n 4. Classication system The pattern classication theory has been a key factor in the development of fault diagnosis methods [21]. Some classication methods for process monitoring use the relationship between a set of patterns and fault types without modeling the internal processes or structure of an explicit way. Nowadays, the ANN is the most popular method. An articial neural network is an information processing paradigm inspired by biological nervous systems. The human learning process may be partially automated with ANNs, which can be congured for a specic application, such as pattern recognition or data classication, through a learning process. An articial neuron consists of a set of connections, which receive and transfer information, and a net function that collects this information (normally as a linear combination weights inputs bias) and sends it to the transfer function, which processes it and produces an output. The basic parts of the general neuron are illustrated in Fig. 7. (5)

ARTICLE IN PRESS
294 n et al. / Mechanical Systems and Signal Processing 24 (2010) 289299 C. Castejo

Fig. 6. Wavelet decomposition at the fth level. Normal Bearing (A30602) and Inner race fault Bearing (I10601) Upper, the original signals; a5-the 5th level approximation (078.125 Hz), d1d5: the ve details, d1 (12502500 Hz), d2 (6251250 Hz), d3 (312.5625 Hz), d4 (156.25312.5 Hz) and d5 (78.125156.25 Hz).

Fig. 7. Articial Neuron diagram.

There are two main phases in the ANNs application: the learning or training phase and the testing phase. The learning phase is critical because it determines the type of future tasks the network will be able to solve. Once the network is trained, the testing phase is followed, whereby the representative features of the inputs are processed. After calculating the weights of the network, the values of the last layer of neurons are compared with the desired output to verify the suitability of the design.

4.1. Multilayer perceptron An multilayer perceptron (MLP) network is designed as each output neuron (call perceptron) is connected to all the previous layer neurons and there is no connection between neurons of the same layer [22]. Perceptron computes a single output from multiple real-valued inputs by forming a linear combination according to its input weights. An example of an MLP Network is shown in Fig. 8. An MLP is a supervised network, and requires a desired response to be trained. The net learn how to transform input data into a desired response, so it is widely used for pattern classication. An MLP can approximate virtually any input/ output map. It has been shown to approximate the performance of optimal statistical classiers in difcult problems. Most ANN applications involve MLPs. A general training behavior is given in Fig. 9; it shows that training and validation data can be separated into distinct sets. Data used for training purposes are partitioned into 20 subsets in order to perform the training process, and the results are given for the test set and its successful rate, classifying most of the 20 signal parts.

ARTICLE IN PRESS
n et al. / Mechanical Systems and Signal Processing 24 (2010) 289299 C. Castejo 295

Fig. 8. Typical MLP structure.

Fig. 9. MLP training performance.

4.2. Neural Networks for classication By using the characteristic vectors (Vcf n of Eq. (5)) as inputs in an ANN, it is possible to classify the signals in its condition. In this manuscript, multilayer perceptron has been chosen as the supervised network for classication purposes. Seventy-ve percent (75%) of all signals obtained from the experimental platform are used for net training, and leave the rest of the signals for testing. In order to select a good neural network conguration, there are several factors to take into consideration. The major points of interest regarding the ANN topology selection are related to

(1) Network design points such as determining the input and output variables, the number of input and output nodes to be used, the number of hidden layers in the neural network, the number of hidden nodes in each hidden layer, and the size of the training data and testing data sets. (2) Network training considerations such as initialing the network weights, choosing proper training parameter values (such as the learning rate), and selecting the training termination criteria. (3) Practical considerations such as network accuracy, network robustness, and implementation feasibility. Even though selecting these parameters is partly a trial-and-error process, there are some guidelines that can be used in choosing these values. For readers who are interested in a detailed description, please refer to [22,23]. In our case, the network architecture consist of the input layer, which contains 18 neurons (the size of the characteristic vector); there is only one hidden layer with 10, 20 and 30 neurons, in order to verify the ANNs behavior; and the output layer has four neurons, one for each condition to classify. The net function is commonly linear, but the transfer function used was the hyperbolic tangent sigmoid. This structure is the most popular one and has been proven to be able to learn complicated functions [23]. The choice of training parameters is sometimes critical to the success of the neural network training process. Unfortunately, the choice of these values is generally problem-dependent. There is no generic formula that can be used to choose these parameter values. Nevertheless, some guidelines, which are described below, can be followed

ARTICLE IN PRESS
296 n et al. / Mechanical Systems and Signal Processing 24 (2010) 289299 C. Castejo

as an initial trial. After a few trials, the neural network designer should have enough experience to set the appropriate parameters for any given problem. In relation to the training design, the targets for supervised learning are the following: Normal conditions: 1 1 1 1. Inner race faults: 1 1 1 1. Outer race faults: 1 1 1 1. Ball fault: 1 1 1 1. The number of hidden layers and hidden nodes affect both the network accuracy and the time required to train the neural network. In this paper, we limit ourselves to a three-layer net, because this structure has been proven to be able to learn arbitrarily complicated continuous functions [24]. In order to perform an objective comparison of various neural networks, it is of interest to set the number of hidden nodes at 10, 20 and 30 and for different tests. It is a common practice to choose a set of training data and a set of testing data that are statistically signicant to represent the system under consideration. The set of training data is used to train the neural network, and it includes validation, while the set of testing data set is used to test the accuracy of the network after the network has been successfully trained. We will use the preprocessed sets extracted from the experimental setup. The normalization of the input values will increase the numerical stability of the neural network processing, while the normalization of the output values is necessary because of the characteristics of the activation function that is used in processing units of the neural network; for these reasons, inputs and outputs are normalized between 0 1 values. One of the critical parameters in neural network training is the speed of convergence, which is determined by the learning coefcient. In general, it is desirable to have fast learning, but not so fast as to cause instability of learning iterations. Different measures can be used to train and monitor the training process. In this manuscript, the two-norm error measure is used to train and monitor the performance of the neural network, as is described in r P 1X ^ xp ; w yxp 2 ^ x; w yxk2 y (6) E ky P p 1 ^ x; w is the network output and yxp is the actual system output with the same input xp given, and w is where y the vector of the current network weights. Popular criteria used to terminate network training are: (a) sufciently small mean-square training error and (b) sufciently small changes in training error. How sufciently small is usually up to the network designer and based on the desired level of accuracy of the neural network. In this work, network training is nished when either the root mean-square error of the training data set or the change in network error is less than 0.005. Information is stored in a feedforward net in its weights. The more weights a network has, the more information it can store. The number of weights is a function of the number of hidden nodes for a three-layer feedforward net. Thus, the more hidden nodes a network has, the more information the net can store. Furthermore, the network is able to learn faster [23]. The initial weights of the neural network play a signicant role in the convergence of the training method. Without a priori information about the nal network weights, it is common practice to initialize all weights with random numbers of small absolute values between 0:1; 0:1. In linear vector quantization and derived techniques, it is usually required to renormalize the weights at every training epoch. The training process also involves a base algorithm for performing the training iterations. The updated error backpropagation with momentum method shown in Eq. (7) was used to update the weights in a network. The learning rate used in this case was 0.8.
m Wm i;j t 1 W i;j t Z

^ @F @W m i;j

(7)

where W m i;j indicates the weight value for the ith neuron, for the training sample j, in the hidden layer m, Z is the learning ^ is an MSE approximation. rate, and F The stopping criteria are based on the minimum error reached, in this case 5%. The network training is also limited to 25 000 epochs and the validation dataset may affect the training, with a maximum of 1000 iteration failed. The initial weights of the neural network play a signicant role in the training convergence. Without a priori information about the nal network weights, it is common practice to initialize all weights with random numbers of small absolute values between 0:1; 0:1. 5. Results and discussion In this section, results of the complete monitoring process are presented. As we commented before, data collected from the lab bench was processed with WP in order to obtain the characteristic vector. In Fig. 10, the experimental process is presented. Squares represent the methodology and bubbles the experimental results obtained. In Table 1, design parameters of the MLP network trained for classication is presented.

ARTICLE IN PRESS
n et al. / Mechanical Systems and Signal Processing 24 (2010) 289299 C. Castejo 297

Fig. 10. Summary of the methodology and results.

Table 1 MLP network design parameters for training. Input value normalization Network input distribution Training Validation Learning type Learning rate (Z) Weights initialization Stopped criteria Max. iterations Limit MSE Minimum gradient Maximum number of fail 1; 1 range 70% 25% Supervised 0.8 Random 25 000 10% 1 1015 1000

Fig. 11. (a) Normal bearing classication. (b) Inner race fault classication. (c) Ball fault classication. (d) Outer race fault classication.

The network has been trained with three different numbers of neurons (10, 20 and 30) in the hidden layer, in order to obtain the best results and to study the inuence of the number of processing units in the hidden layer in the training and classication process.

ARTICLE IN PRESS
298 n et al. / Mechanical Systems and Signal Processing 24 (2010) 289299 C. Castejo

Table 2 Success rates at different speeds. RPMs 600 Normal 100 100 100 100 100 100 100 100 100 Inner race default 33 43 61 69 82 86 71 80 90 Ball default 100 100 100 100 100 100 100 100 100 Outer race default 61 84 82 16 25 41 0 8 18 Hidden neurons 10 20 30 10 20 30 10 20 30 Network accuracy 73.47 81.63 85.71 71.43 76.53 81.63 67.86 71.94 77.04

1200

1800

In Fig. 11, the results of classication of different bearing conditions for different numbers of hidden neurons are presented. This study lets us determine the best network conguration for a specic industrial problem (that will depend on the machine complexity and the rotor speed). Normal condition and Ball fault condition are easily diagnosed and they do not show any dependence with the number of hidden neurons, which indicate the ANN complexity level. Outer race faults are diagnosed with less success rate, and it is shown that this network works better at a higher number of hidden neurons. Finally, the inner race fault cannot be classied at a good success rate, but improves its behavior at a higher complexity. Nevertheless, better results of this classication have been found for other rotor speeds. Additional tests were performed at higher speeds (1200 and 1800 RPM), reporting an improvement in the fault bearing success rate but decreasing the outer race fault success rate, as illustrated in Table 2. As can be seen in Table 2, the best result of classication can be obtained when the rotor speed is 600 RPM and the number of hidden neurons is 30. The network accuracy in this case is around 85%. Related to other works presented in the literature which show better results, we must say that, due to the experimental setup, the contact ellipse between ball and race is smaller than the contact ellipse in our work. Therefore, it is reasonable to think that the results obtained in the manuscript are of interest and they are not comparable with other manuscripts which do not consider the incipient fault bearing condition under radial load. 6. Conclusions In this manuscript, an automatic fault classication technique based on MRA and neural networks has been developed. The difculty in classifying bearing conditions from data provided by a real machine is also stated in this work. The neural network MLP can be used to classify the bearing condition in a very incipient stage with a success rate of up to 80%. DWT can reliably extract features for classication tasks, allowing early fault detection at incipient levels. Since the basis functions used on the WT are compactly supported, wavelets have good properties of energy concentration. Furthermore, the proposed absolute classication method lets us take a measure of the accuracy in well classied bearings and not in classied signals. The method values the diagnosis repetition and lets us continuously evaluate the dynamic state of the system, knowing the change of fault state in advance. The method proposed lets us automate the complete maintenance process in a real industry.

Acknowledgments The authors wish to thank the Spanish government for nance provided through the MCYT project DPI-2006-15443C02-02 and the Mechanical Department of UNED for its assistance with the experimental part. References
[1] Q. Hu, Z. He, Z. Zhang, Y. Zi, Fault diagnosis of rotating machinery based on improved wavelet package transform and SVMs ensemble, Mechanical System and Signal Processing 21 (2007) 688705. [2] K.L.X. Lou, Bearing fault diagnosis based on wavelet transform and fuzzy inference, Mechanical Systems and Signal Processing 18 (2004) 10771095. [3] T.A. Harris, Rolling Bearing Analysis, fourth ed., A Wiley-Interscience Publication, USA, 2001. [4] G. Wang, Z. Luo, X. Qin, Y. Leng, T. Wang, Fault identication and classication of rolling element bearing based on time-varying autoregressive spectrum, Mechanical Systems and Signal Processing 21 (2007) 934947. [5] V. Rai, A. Mohanty, Bearing fault diagnosis using FFT of intrinsic mode functions in HilbertHuang transform, Mechanical Systems and Signal Processing 21 (2007) 26072615. [6] Q. Du, S. Yang, Application of the EMD method in the vibration analysis of ball bearings, Mechanical Systems and Signal Processing 21 (2007) 26342644.

ARTICLE IN PRESS
n et al. / Mechanical Systems and Signal Processing 24 (2010) 289299 C. Castejo 299

[7] Y. Lei, Z. He, Y. Zi, Q. Hu, Fault diagnosis of rotating machinery based on multiple ans combination with GAs, Mechanical Systems and Signal Processing 21 (2007) 22802294. [8] B. Taoa, L. Zhub, H. Dinga, Y. Xionga, An alternative time-domain index for condition monitoring of rolling element bearings. A comparison study, Reliability Engineering and System Safety 92 (2007) 660670. [9] S. Mallat, A Wavelet Tour of Signal Processing, Academic Press, New York, 1998. [10] P.M.W.J. Wang, Application of the wavelet transform to Garvox vibration analysis, American Society of Mechanical Engineers, Petroleum division, 1993, pp. 1320. [11] Z. Peng, Application of the wavelet transform in machine condition monitoring and fault diagnostics: a review with bibliography, Mechanical Systems and Signal Processing 18 (2004) 199221. [12] B. Liu, S. Ling, Machinery diagnostic based on wavelet packets, Journal of Vibration and Control 3 (1997) 517. [13] B. Chen, X. Wang, S. Yang, C. McGreavy, Application of wavelets and neural networks to diagnostic system development, 1, feature extraction, Computers and Chemical Engineering 23 (1999) 899906. [14] A. Rainch, A. Cinar, Statistical process monitoring and disturbance diagnosis in multivariate continuous processes, A.I.Ch.E. Journal 42 (1996) 9951009. [15] T. Sorsa, H. Koivo, Application of articial neural networks in process fault diagnosis, Automatica 29 (1993) 843849. [16] W. Staszewski, Vibration data compression with optimal wavelet coefcients genetic algorithms in engineering systems, in: Innovations and Applications, vol. 446, 1997, pp. 186189. [17] M. Chow, R. Sharpe, J. Hung, On the application and design of articial neural netwoeks for motor fault detection-part I, IEEE Transactions on industrial electronics 40 (1993) 181188. [18] M. Chow, R. Sharpe, J. Hung, On the application and design of articial neural netwoeks for motor fault detection-part II, IEEE Transactions on Industrial Electronics 40 (1993) 189196. [19] C. Scheffer, P. Girdhar, Practical Machinery Vibration and Predictive Maintenance, Elsevier, Amsterdam, 2004. a-Prada, C. Castejo n, O. Lara, Incipient bearing fault diagnosis using DWT for feature extraction, in: 12th IFToMM World Congress, Besanc [20] J. Garc - on, France, 2007. n, J. Garc a-Prada, Bearing fault diagnosis based on neural network classication and wavelet transform, WSEAS Transaction on [21] O. Lara, C. Castejo Signal Processing 2 (10) (2006) 13711378. [22] M. Godoy, M. Meireles, P. Almeida, A comprehensive review for industrial applicability of articial neural networks, IEEE Transactions on Industrial Electronics 50 (2003) 585601. [23] M. Chow, Methodologies of using Neural Network and Fuzzy-logic Technologies for Motor Incipient Fault Detection, World Scientic, Singapore, 1997. [24] G. Cybenko, Approximation by superpositions of a sigmoidal function, Mathematics of Control, Signals, and Systems (MCSS) 2 (4) (1989) 303314.