ARTICLE IN PRESS

Mechanical Systems and Signal Processing 23 (2009) 712723

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

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

Normalized wavelet packets quantiers for condition monitoring

Yanhui Feng , Fernando S. Schlindwein
Department of Engineering, University of Leicester, Leicester LE1 7RH, UK

a r t i c l e i n f o
Article history: Received 7 October 2007 Received in revised form 17 July 2008 Accepted 17 July 2008 Available online 24 July 2008 Keywords: Condition monitoring Wavelet packets quantier Bearing fault Contamination fault Wavelet entropy Acoustic Emission

abstract
Normalized wavelet packets quantiers are proposed and studied as a new tool for condition monitoring. The new quantiers construct a complete quantitative timefrequency analysis: the Wavelet packets relative energy measures the normalized energy of the wavelet packets node; the Total wavelet packets entropy measures how the normalized energies of the wavelet packets nodes are distributed in the frequency domain; the Wavelet packets node entropy describes the uncertainty of the normalized coefcients of the wavelet packets node. Unlike the feature extraction methods directly using the amplitude of wavelet coefcients, the new quantiers are derived from probability distributions and are more robust in diagnostic applications. By applying these quantiers to Acoustic Emission signals from faulty bearings of rotating machines, our study shows that both localized defects and advanced contamination faults can be successfully detected and diagnosed if the appropriate quantier is chosen. The Bayesian classier is used to quantitatively analyse and evaluate the performance of the proposed quantiers. We also show that reducing the Daubechies wavelet order or the length of the segment will deteriorate the performance of the quantiers. A twodimensional diagnostic scheme can also help to improve the diagnostic performance but the improvements are only signicant when using lower wavelet orders. Crown Copyright & 2008 Published by Elsevier Ltd. All rights reserved.

1. Introduction Failure of rolling element bearings will cause machine malfunction and may quickly lead to catastrophic failure of the machinery if no early maintenance is undertaken. Premature bearing failures can be caused by a large number of factors which can be systematically divided into four groups for study [1]: design, production technology, operation and change of condition (DPTOCC). The design factors are taken by a designer to give the background for the bearing production, including all the details connected with the shape of the structure, bearing selection and imperfection limits. The factors connected with the production are called production technology factors, which may or may not full the design factors. The factors connected with bearing operation are named operation factors, including load, rotation speed and environment condition (dustiness, moisture and so on). The change of condition comes from a fault occurring during operation of bearing, such as localized defect, distributed defect and contamination fault. In this paper, the localized defect and contamination fault will be studied. These factors will lead to DPTOCC inferring diagnostic information of a bearing system. The development of modern signal processing techniques for bearing faults detection using vibration signals is given in [26]. Acoustic Emission (AE) (i.e. structure-borne ultrasound) measurements are a subject of great recent interest because of their improved sensitivity for some advanced faults compared to the traditional vibration measurements [7].

Corresponding author.

E-mail address: yf12@le.ac.uk (Y. Feng). 0888-3270/$ - see front matter Crown Copyright & 2008 Published by Elsevier Ltd. All rights reserved. doi:10.1016/j.ymssp.2008.07.002

ARTICLE IN PRESS
Y. Feng, F.S. Schlindwein / Mechanical Systems and Signal Processing 23 (2009) 712723 713

Localized bearing defects include cracks, pits and spalls on the rolling surfaces. The vibration signals caused by localized defects can be modelled as a second-order cyclostationary process [8]. Many methods have been proposed and studied for localized defects detection, such as the high-frequency resonance technique [9] and cyclostationary signal analysis [6]. The frequency of the periodic impacts, termed as the bearing Characteristic Defect Frequency (CDF), can be estimated by equations of shaft speed, bearing geometry and defect location [3]. Signal processing methods on frequency domain for localized defects detection are used for trying to nd out if signicant power exists at the CDFs. Surface-initiated damage due to contamination is one of the main reasons for early bearing failure. At an early stage, the contamination fault does not produce any harmonics of shaft speed frequency at low frequencies and there is often no obvious CDF being modulated at high frequencies. The Wavelet Transform (WT) is the time-scale analysis method with octave decomposition and the advantage of wavelet analysis lies in detecting transient changes [10,11]. The applications of the WT for condition monitoring have been developed quickly over the last 10 years [12]. As an extension of Discrete Wavelet Transform (DWT), Discrete Wavelet Packets Transform (DWPT) is a powerful and versatile quantitative joint timefrequency analysis method for condition monitoring. This paper proposes and studies normalized wavelet packets quantiers as a new tool for the detection and diagnosis of localized bearing defect and contamination fault. Unlike the feature extraction methods, which use the amplitude of wavelet coefcients, these new quantiers are derived from probability distributions and are more robust for diagnostic applications. This paper is organized as follows: Section 2 provides an overview of the methodology. Section 3 describes the experiment set-up and data acquisition. Section 4 shows the results and performance evaluation of applying the new quantiers to AE signals from faulty bearings of rotating machines. We discuss the results in Section 5 and nally report conclusions in Section 6. 2. Methodology 2.1. Discrete wavelet packets transform The DWPT yields timefrequency decomposition by simply decomposing the details of DWT coefcients into ner and ner dyadic frequency bands [10]. By exibly choosing the nodes from DWPT, we can have better insight of the signal timefrequency structure. The basis of DWPT can be exibly chosen for practical implementation considerations; some packet nodes which concentrate the contaminating noises can be easily eliminated; the computational time for DWPT is fast because it is supported by the fast lter bank algorithm. The DWPT generates a wavelet packets table or wavelet packets tree. The wavelet node (j, n) corresponds to the vector Wj,n of wavelet packets coefcients where j 1,y,J0 is the depth of the node and n 0,y,2j1 is the number of nodes that are on its left at the same depth. The collection of nodes forming the indices of wavelet packets table will be denoted by T{(j, n): j 1,y,J0; n 0,y,2j1}. The coefcient Wj,n,t of node (j, n) at time t can be written in terms of a ltering of signal S with appropriate downsampling [11]:
Lj 1

W j;n;t

X
l0

uj;n;l S2j t11l modulo N ;

t 0; 1; . . . ; N j 1,

(1)

where Nj is the length of coefcients vector Wj,n at level j and Nj N/j, Lj is the lter width at j level Lj (2j1)(L1)+1 and uj,n,l is the lter for node (j, n). JWj,nJ2 can be interpreted as the energy portion in the frequency band [fs n/2j+1, fs (n+1)/2j+1], that is
N j 1

kWj;n k2 Ej;n

X
t 0

Wj;n;t .

(2)

2.2. New wavelet packets quantiers ^ ) preserve the energy of signal S, Since the DWPT is an orthonormal transform, the terminal wavelet packets nodes (^ j; n that is kSk2
J0 X 2 X ^ ^ 0 j 1 n
^ j1

2 kW^ ^k : j; n

(3)

We can eliminate terminal nodes for denoising purpose (e.g. after wavelet packets composition of AE signals, the energy of the nodes corresponding to low frequencies noise will be set to zero). In this case, the total energy of the remaining terminal nodes collection C is X Etot E^ (4) ^. j; n
^ 2C ^ j;n

ARTICLE IN PRESS
714 Y. Feng, F.S. Schlindwein / Mechanical Systems and Signal Processing 23 (2009) 712723

We therefore obtain the normalized probability distribution {p^ ^ t } in the time domain, over the energy of the selected j;n ^ ): terminal node (^ j; n
2 p^ ^ t jW ^ ^ ;t j =E^ ^; j;n j; n j; n

t 0; 1; . . . ; Nj 1.

(5)

: S

Also, we have the normalized energy distribution {p^ ^ f } in the frequency domain, over the energy of the denoised signal j;n p^ ^ f E^ ^ =Etot j;n j; n (6)

which is the Relative wavelet packets energy. Clearly, {p^ ^ t } and {p^ ^ f } are the probability distributions in timefrequency j;n j;n plane where each of them represents density, such that
N^ 1 j

X
t 0

p^ ^ t j;n

X
^ 2C ^ j;n

p^ ^ f 1. j; n

(7)

Entropy is a very powerful concept in measuring the uncertainty (i.e. the degree of disorder) associated with a random variable. The concept plays an important role in statistical mechanics, information theory, communications, signal processing, data mining and machine learning. The Shannon entropy [13] is the most popular quantier of entropy. Here we dene another two quantiers based on the Shannon entropy. These are the Wavelet packets node entropy
N^ 1 j

Q^ ^ t j;n

X
t 0

p^ ^ t log2 p^ ^ t j;n j;n

(8)

and the Total wavelet packets entropy X p^ Q f ^ f log2 p^ ^ f . j; n j; n

^ 2C ^ j;n

(9)

These two quantiers can provide very useful information about the dynamics of the signal. The rst quantier describes ^ ^ how disordered the normalized coefcients {p^ ^ t } are in a specic node (j; n). If the process shows very ordered behaviour j;n in a node, it tends to concentrate into a few large coefcients and have strong uctuation in the time domain. The value of the quantier will then be near zero. One the other hand, the value of the quantier will be larger if the process is more disordered. In this case, the coefcients will distribute more evenly, without particular strong individual values. The contributions from all the normalized coefcients inside the node are almost equal and consequently, the sum of these contributions takes a larger value. The wavelet packets quantiers were rst studied by Rosso et al. [1416] to characterize brain electrical signals. Total wavelet packets entropy describes the degree of the disorder of signals from the view of frequency energy distribution. Signals with a narrow band spectrum can be viewed as a typical example of ordered frequency energy distribution. In this case, the total energy of the signal is concentrated into a terminal node (i.e. the quantity of Relative wavelet packets energy of this node is very large) and the energies in other nodes are very small. The value of the quantier Total wavelet packets entropy will then be relatively small. On the other hand, if the energies are rather equally distributed to the terminal nodes, the whole frequency energy distribution tends to be more disordered and the value of the Total wavelet packets entropy therefore will be larger. The new quantiers construct a complete quantitative timefrequency analysis based on probability distribution: the Wavelet packets relative energy measures the normalized energy of the wavelet packets node; the Total wavelet packets entropy measures how the normalized energies of the wavelet packets nodes are distributed in the frequency domain; the Wavelet packets node entropy describes the degree of disorder of the normalized coefcients of the wavelet packets node. 2.3. Decomposition conguration We divide AE signals from different bearing conditions into many N-point segments S, then decompose all these segments at 2 levels using DWPT and produce 4 terminal wavelet packets nodes (2,0), (2,1), (2,2), (2,3) corresponding to frequency bands (0, 10 kHz), (10, 20 kHz), (20, 30 kHz) and (30, 40 kHz), respectively. The last three bands are labelled as A, B and C (see Fig. 1). For AE signals mainly focusing on the very high frequency, the band (0, 10 kHz) is not used in the analysis and is discarded. All the three remaining nodes are analysed using the above proposed quantiers. Specically, the is total energy of denoised signal S X Etot Ek , (10)
kA;B;C

where Ek t 0 Wk,t for k A, B, C band. The Relative wavelet packets energy of band k is pk f Ek =Et , (11)

P N/41

ARTICLE IN PRESS
Y. Feng, F.S. Schlindwein / Mechanical Systems and Signal Processing 23 (2009) 712723 715

T (1,0)

T (1,1)

T (2,0)

T (2,1) A

T (2,2) B

T (2,3) C

Fig. 1. Two-level DWPT on the AE signals (frequency ordering).

the Wavelet packets node entropy in the band k is Q k t

NX =41 t 0

pk t log2 pk t ,

(12)

and the Total wavelet packets entropy of the segment is X Q f pk f log2 pk f .

kA;B;C

(13)

2.4. Performance evaluation method The problem of bearing fault detection and diagnosis can be systematically studied in the context of pattern recognition [17]. We will evaluate the performance of the proposed quantiers, also known as the features in pattern recognition, by comparing their classication error probability. The adopted classication method based on Bayes decision theory is optimal with respect to minimizing the classication error probability or the total risk. Consider a classication task of M classes, oi, i 1,2,y,M, the a posteriori probability of an unknown pattern, which is represented by a feature vector X, is denoted by P(oi|X). It represents the probability that the unknown pattern belongs to the respective class oi, given the corresponding feature vector takes X. For XARi, the posteriori probability P(oi|X) is the correct classication probability; otherwise, it is the classication error probability. The risk associated with the class ok is rk
M X i 1

lki

Z
X 2Ri

pXjok dX:

(14)

The total risk for the Bayesian classication is r

M X k1

r k P ok

M X M X k1 i1

lki

Z
X 2R i

pXjok Pok dX;

(15)

and this can be written in the form of posteriori probability as r

M X M X k1 i1

lki Pok jX 2 Ri .

(16)

Probability density function P(X|oi) describes the distribution of the feature vectors in each class. The integral in the region Ri of the class oi represents the classication error probability of ok caused by the misclassication, i.e. the values of feature belonging to class ok being misclassied as belonging to class oi. The penalty term lki is zero for k i because no penalty is given to correct classication. When there is no practical risk evaluation available, we assume lki is 1 for any k6i. That is, any misclassication has equal penalty. The above total risk becomes the total classication error probability E
M X M X k1 i1

P ok jX 2 Ri .

(17)

If the probability densities of different classes with respect to a quantier are less overlapped than other quantiers, they have a greater capability to discriminate and diagnose different bearing conditions.

ARTICLE IN PRESS
716 Y. Feng, F.S. Schlindwein / Mechanical Systems and Signal Processing 23 (2009) 712723

3. Experimental setup and data acquisition A ve-stage Roots and Claw dry vacuum pump with empty load was used as test bed. A bearing with three conditions was mounted at the high vacuum side of pump. These bearing conditions were: fault free, localized indent defect on inner race (see Figs. A1 and A2 in Appendix A) and contamination fault. The speed of pump was set at 105 Hz (6300 rev min1) and the inlet pressure was set at 0 mbar. An AE transducer (PAC R3a) was rmly held at the surface of the pump house near the high vacuum end in the radial direction. The acquired signals were amplied with the gain of 1000 and ltered by an anti-aliasing lter. Signals were digitized by a 16-bit NI Analogue to Digital Converter with the sampling rate 200 kHz and stored in a computer. The signals were downsampled to 80 kHz ofine and further processed in Matlab.

4. Results The acquired AE signals from different bearing conditions (fault free, contamination fault and localized defect) are divided into many L-point segments without overlap and subsequently decomposed by the 2-level DWPT using the Daubechies wavelet bases. The produced wavelet packets nodes are denoised and the A, B, C bands are obtained as mentioned in Section 2.3. These nodes are then analysed using the proposed quantiers. Firstly, signals are divided into 292 sets of 2048-point segments and decomposed by the 2-level DWPT using db10 as wavelet basis. The histograms of the quantities of the quantier Relative wavelet packets energy of A, B and C bands for three different bearing conditions are shown in Figs. 24, respectively. The distributions are well tted with Gaussian
60 50 Number of Data Blocks 40 30 20 10 0 0.60 Fault free Contaminated fault Localized defect

0.65 0.70 0.75 0.80 0.85 0.90 Relative wavelet packets energy of A band

0.95

Fig. 2. The histogram of Relative wavelet packets energy of A band for three classes (L 2048, db10).

80 70 Number of Data Blocks 60 50 40 30 20 10 0 0.00 Fault free Contaminated fault Localized inner race defect

0.05 0.10 0.15 0.20 0.25 0.30 0.35 Relative wavelet packets energy of B band

0.40

Fig. 3. The histogram of Relative wavelet packets energy of B band for three classes (L 2048, db10).

ARTICLE IN PRESS
Y. Feng, F.S. Schlindwein / Mechanical Systems and Signal Processing 23 (2009) 712723 717

70 60 Number of Data Blocks 50 40 30 20 10 0 0.00

Fault free Contaminated fault Localized inner race defect

0.02 0.04 0.06 Relative wavelet packets energy of C band

0.08

Fig. 4. The histogram of Relative wavelet packets energy of C band for three classes (L 2048, db10).

25

Number of Data Blocks

20

Fault free Contaminated fault Localized defect

15

10

0 0.3

0.4

0.5 0.6 0.7 0.8 0.9 Total wavelet packets entropy

1.0

1.1

Fig. 5. The histogram of Total wavelet packets entropy for three classes (L 2048, db10).

distribution (prole in solid curves), likewise for the following gures. In Fig. 1, most energy of the signals is concentrated in the A band (10, 20 kHz): over 65% of the energy for contamination fault, over 80% for fault free and over 85% for localized defect. The energy of B band (20, 30 kHz) is the second largest. The energy for contamination fault is over 18%, for fault free, about 1018% and for localized defect, no more than 8%; see Fig. 3. We should also notice that the three classes here are well separated except a slight overlap between fault-free and contamination fault. Compared to A and B band, C band (30, 40 kHz) contains the lowest energies; see Fig. 4. Only 28% energies are for localized defect and less than 2% energies for other two classes. When contamination fault happens, it shows lowest energy portion in band A compared to the other two bearing conditions. However, relative high energies trend to spread to band B and almost all the energies attenuate in band C. In contrast, localized defect has very different characteristics. It shows relatively high energies in A and C bands, however lower energy in band B. In all the three bands, the energy probability distributions for fault-free condition are located in the middle. To further investigate these energy distributions in three bands, we plot the histogram of the quantities of the Total wavelet packets entropy for three bearing conditions (see Fig. 5). This quantier describes the degree of the disorder of signals from the view of frequency energy distribution. Narrowband-like signals are more ordered than wideband-like signals. If the total energies of signals concentrate in one band, the quantities will then have relatively small values. From the previous graph, we can see that localized defects produce more narrowband-like signals, where over 85% energies concentrate in A band, while bands B and C have together no more than 15% of the energy. Contamination fault signal produces more broadband-like signals, where 6580% of the energy is in band A and 1835% of the energy in band B.

ARTICLE IN PRESS
718 Y. Feng, F.S. Schlindwein / Mechanical Systems and Signal Processing 23 (2009) 712723

Table 1 Classication error probability using different quantiers (L 2048 and db10) Error Fault free Contamination Loc. defect Total error F1 0.048 0.003 0.031 0.081 F2 0.003 0.003 0 0.006 F3 0.182 0.092 0 0.273 F4 0.151 0.010 0.140 0.300 F5 0.699 0.387 0.397 1.482 F6 0.702 0.469 0.188 1.359 F7 0.175 0.592 0 0.768

NB: since the total error adds false positives and false negatives it might be larger than 1.

110 100 Number of Data Blocks 90 80 70 60 50 40 30 20 10 0 4 5 6 7 8 Wavelet packet node entropy in band C
Fig. 6. The histogram of Wavelet packets node entropy in band C for three classes (L 2048, db10).

Fault free Contaminated fault Localized inner race defect

Broadband-like signals are more disordered than narrowband-like signals. The energies of wideband-like signals tend to spread the energies over the bands and the quantier therefore takes larger values. According to the above analysis, we may expect the relationship of the quantities of the quantier Total wavelet packets entropy for three bearing conditions to be: contamination fault4fault free4localized defect. Fig. 5 shows a good agreement with this expectation. We again notice that the overlaps between contamination fault and the other two conditions are very small, which implies that this quantier is good at detecting contamination faults. The quantitative analysis for calculating the overlaps will be given in Table 1. As mentioned in Section 2.2, the quantier Wavelet packets node entropy is used to describe how disordered the normalized coefcients are in a frequency band. The quantities of the quantier will take small values if coefcients show ordered behaviour, and vice versa. Fig. 6 shows the histogram of the quantities of the quantier Wavelet packets node entropy in band C. In Fig. 6, the quantities for the localized defect are much lower than the other two, which implies that the coefcients might be very ordered and the processes concentrate energy in a few large coefcients and have stronger peaks uctuation. Fig. 7 displays the normalized coefcients in band C (from top: localized defect, contamination, fault free) and shows good agreement with the above analysis. For fault-free and contamination fault, the processes are relatively disordered and their energies are distributed more evenly to many coefcients. Consequently, the quantities for them take larger values. The overlaps between localized defect and the other two conditions are very small, which implies this quantier is good at detecting localized defects. To quantitatively evaluate the performance of the proposed quantiers, we use the Bayesian classier to compute the classication error probability, where the probability density functions are estimated by the Gaussian distributions. The classier is trained with 30% of data and tested with the remainder. In the following context, the quantier Relative wavelet packets energies of A, B, C bands are denoted as F1, F2 and F3; the quantier Total wavelet packets entropy is denoted as F4; the quantier Wavelet packets node entropy of A, B, C bands is denoted as F5, F6 and F7. Table 1 shows the classication error probability using the quantiers F1F7. The rst row in Table 1 is the classication error probability for the fault-free condition; that is, the patterns belonging to fault free are misclassied to the other two classes. The second row and the third row are the classication error probability for contamination fault and localized defect, respectively. The fourth row is the total classication error probability from the sums of all three types of classication error probability.

ARTICLE IN PRESS
Y. Feng, F.S. Schlindwein / Mechanical Systems and Signal Processing 23 (2009) 712723 719

0.3 0.2 0.1 0.0 -0.1 -0.2 -0.3 0.3 0.2 0.1 0.0 -0.1 -0.2 -0.3 Contamination fault Localized defect

0.3 0.2 0.1 0.0 -0.1 -0.2 -0.3 0 Fault free 100 200 300 400 500

Fig. 7. The normalized wavelet packets coefcients in band C (L 2048, db10).

Signicant low classication error probability (less than 5%) is highlighted in bold. The best diagnosis performance is from F2, where total error probability is only 0.6%. The second best performance is from F1, where total error probability is 8.1%. We can explain the results from the corresponding Figs. 2 and 3, both showing small overlaps among different classes. In Fig. 3, all three classes are well separated except slight overlaps between fault free and contamination fault. The results in Table 1 show the small overlaps cause 0.3% classication error probability for each class. Because the distribution of localized defect has no overlap with the other two classes, the classication error probability for it is zero. In Fig. 2, more overlaps between localized defect and fault-free classes appear, which cause the increasing classication error probability 4.8% and 3.1% for each class. The rest of the ve quantiers have not attained satisfactory results, with their total error probabilities over 25%. Although their total performances are not satisfying, these quantiers have very low classication error probability for the classes relevant to some faults (e.g. F3 and F7 for localized defect, F4 for contamination fault), which means they are quite sensitive to specic types of bearing fault. In some situations, a twodimensional diagnostic scheme utilizing these quantiers can help to reduce the total classication error probability. For implementation, wavelet order, length of segment and dimension(s) are often important considerations. They will be studied as follows. In the above results, we used Daubechies wavelet dbN with large order N 10 as wavelet packets basis. Now we investigate how the change of order inuences the performance of the quantiers. Fig. 8 shows the total classication error probability of the proposed quantiers when using different Daubechies wavelet orders (N 1, 2,y,10). The graph shows F5 and F6 both have the worst performances with unacceptable classication error probability. F4 and F7 also perform badly with over 30% classication error probability for all the wavelet orders. F3 only performs well when using db1 (haar) basis. F2 has the best performance for the larger wavelet orders but the performance deteriorates when wavelet order goes lower. F1 is the second best quantier overall, but the classication error probability also increases when wavelet order decreases. Next, we investigate how the length of segment inuences the performance of the quantiers. We divide the signals into segments of different sizes still keeping 292 sets of segments for Bayesian classication. Fig. 9 displays the total

ARTICLE IN PRESS
720 Y. Feng, F.S. Schlindwein / Mechanical Systems and Signal Processing 23 (2009) 712723

1.6 1.5 1.4 1.3 1.2 1.1 1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 db10 db9 db8 db7 db6 db5 db4 db3 Daubechies wavelets db2 db1

Fig. 8. Total classication error probability of the quantiers using different Daubechies wavelet orders (N 1, 2,y,10) (L 2048).

Total classification error probability

F1 F2 F3 F4 F5 F6 F7

Total classification error probability

0.30 0.25 0.20 0.15 0.10 0.05 0.00 N = 2048 N = 1024 N = 512 N = 248 N points' data block

0.30 0.25 0.20 0.15 0.10 0.05 0.00

Fig. 9. Total classication error probability using different lengths of segment (db10).

classication error probability when using different lengths of segment for the quantier F2 and the wavelet basis db10. The previous analyses show the quantier F2 and the Daubechies wavelet dbN with large order N 10 can achieve the best performance of all. Fig. 9 shows the total classication error probability signicantly increases when the length of segment decreases. Fig. 8 shows how the performance of the best quantier F2 deteriorates when wavelet order decreases. When using very low order db1 (Haar), the classication error probability increases to 27.6%, where two classesfault free and contamination faultoverlap heavily. Fig. 10 shows the histogram for this situation. A possible approach to solve this problem is to use a two-dimensional diagnostic scheme, i.e. combine a pair of quantiers to diagnose bearing faults. Fig. 11 shows a two-dimensional diagnostic map using two quantiers F2 and F4. The quantier F4 is used to help for discriminating the contamination fault and fault-free cases, where the total classication error probability reduces from 27.6% to 9.6%. Fig. 12 shows the comparison of classication error probability between the best single quantier F2 and the best combination pair of quantiers. It shows two-dimensional diagnostic scheme can help to improve the diagnostic performance. The improvements are more signicant when using lower wavelet orders.

ARTICLE IN PRESS
Y. Feng, F.S. Schlindwein / Mechanical Systems and Signal Processing 23 (2009) 712723 721

40 Number of Data Blocks

Fault free Contaminated fault Localized defect

30

20

10

0 0.2 0.3 0.4 Relative wavelet packets energy of band B

Fig. 10. The histogram of quantier F2 for three classes using Haar wavelet (L 2048); the total classication error probability is 27.6%.

Relative wacelet packet energy of band B (F2)

0.4 0.35 0.3 0.25 0.2 0.15 1.1 1.15 1.2 1.25 1.3 Total wavelet packet entropy (F4) 1.35 1.4
Fault Free Contamination fault Localized defect

Fig. 11. Two-dimensional diagnostic map using two quantiers F2 and F4 (L 2048).

Total classification error probability

0.25 0.20 0.15 0.10 0.05 0.00 haar db2

Best single feature (F2) 2-d best combination pair

db3

db4

Daubechies wavelets
Fig. 12. Comparison of classication error probability (L 2048).

ARTICLE IN PRESS
722 Y. Feng, F.S. Schlindwein / Mechanical Systems and Signal Processing 23 (2009) 712723

5. Discussion The new quantiers derived from probability distributions are either normalized over the total signal energy or the node energy. Here we emphasise the need for normalization for robust diagnostic application. In practice, it is very difcult to compare different bearing conditions using the quantiers without normalization, unless the amplier is carefully tuned to achieve equal amplitudes for different signals. The signal with higher amplitude will produce the wavelet coefcients with higher amplitudes; it thus generates unpredictable effects on the quantiers. For example, the signal from localized defect without normalization will generate much higher amplitude coefcients in band B than we can see in Fig. 3, where the class belonging to the localized defect merges into two other classes. From the viewpoint of implementation, we expect the wavelet basis to have higher vanishing moments (order) N because the higher the vanishing moments, the better representation of signal we will have. Signals can be efciently approximated by fewer non-zero wavelet coefcients. However for the Daubechies orthogonal wavelets, the lter support size has to be at least 2N1 with the vanishing moment N. The implementation of longer support size will increase the cost and slow down the computation. Our study shows the classication error increases when wavelet order N increases for some good quantiers (F1 and F2) with the satisfactory performances overall. Our results also show that the total classication error increases signicantly as the length of segment decreases. A two-dimensional diagnostic scheme using a pair of quantiers can help to reduce the classication error. However, the reductions are only signicant when using lower wavelet order. In summary, an optimal implementation needs to consider how to balance between the increasing classication errors and the support size, length of segment or dimension(s). In this study, we use Bayesian classier to quantitatively evaluate the performance of the proposed quantiers. Other classiers are possible and they deserve further research. We assume lki is 1 for any k6i, i.e. any misclassication has equal penalty. However, this is a naive assumption because in the real world some of the misclassication may imply more serious consequences; therefore, they should be penalized more heavily. When the dry vacuum pump is running with load, the load might have an inuence on the performance of the quantiers. The inuence factor will be investigated in future work.

6. Conclusions In this paper, we propose and study the normalized wavelet packets quantiers as a new tool for condition monitoring. The new quantiers construct a complete quantitative timefrequency analysis: the Wavelet packets relative energy measures the normalized energy of the wavelet packets node; the Total wavelet packets entropy measures how the normalized energies of the wavelet packets nodes are distributing in the frequency domain; the Wavelet packets node entropy describes the uncertainty (i.e. the degree of disorder) of the normalized coefcients of the wavelet packets node. Unlike the feature extraction methods directly using the amplitude of coefcients, these new quantiers are derived from probability distributions and are more robust in diagnostic applications. By applying the technique on AE signals from faulty bearings, our study shows that both localized defects and advanced contamination faults can be successfully detected and diagnosed if the appropriate quantier is chosen. We further use the Bayesian classier to quantitatively analyse and evaluate the performance of the proposed quantiers. We show the optimal quantier for an application can be found out by comparing the total classication error probabilities of quantiers. The optimal quantier for an application can also be found out by comparing the total classication error probabilities of quantiers. In this case study, the Wavelet packets relative energy of band B (20, 30 kHz) is the optimal quantier. We also show the performance of the quantier depends on the wavelet order and the length of segment. Reducing the Daubechies wavelet order or the length of segment will deteriorate the performance of the quantier. We also show that a two-dimensional diagnostic scheme may also help to improve the diagnostic performance and the improvements are more signicant when using lower wavelet orders.

Acknowledgements The authors wish to thank Dr. J. Twiddle for his great assistance with the experiment, and his discussion on the paper. The authors thank BOC Edwards for the supply of the IGX dry vacuum pump and Barden Bearings for providing the bearings for the experimental investigations. Thanks are also given to the reviewers who provided valuable comments to improve the manuscript, particularly on the knowledge of DPTOCC framework and the inuence of load.

Appendix A The bearing conditions were: fault free, localized indent defect on inner race (see Figs. A1 and A2).

ARTICLE IN PRESS
Y. Feng, F.S. Schlindwein / Mechanical Systems and Signal Processing 23 (2009) 712723 723

Fig. A1. Articially made indent on inner race of bearing.

Fig. A2. The seeded indent is approximately 2 mm wide and 2 mm deep.

References
[1] Walter Bartelmus, Mathematical modelling and computer simulations as an aid to gearbox diagnostics, Mech. Syst. Signal Process. 15 (5) (2001) 855871. [2] S. Braun, et al., Mechanical Signature Analysis: Theory and Application, Academic Press Inc. (London) Ltd., London, 1986. [3] Ian Howard, A review of rolling element bearing vibration: detection, diagnosis and prognosis, DSTO-RR-0013, Defence Science and Technology Organization, Canberra (Australia), October 1994. [4] N. Tandon, A. Choudhury, A review of vibration and acoustic measurement methods for the detection of defects in rolling element bearings, Tribol. Int. 32 (1999) 469480. [5] Robert B. Randall, State of the art in monitoring rotating machinery: part 1, Sound Vib. (2004). [6] Robert B. Randall, State of the art in monitoring rotating machinery: part 2, Sound Vib. (2004). [7] D. Mba, Raj B.K.N. Rao, Development of acoustic emission technology for condition monitoring and diagnosis of rotating machines: bearings, pumps, gearboxes, engines and rotating structures, Shock Vib. Digest (2006). [8] J. Antoni, R.B. Randall, A stochastic model for simulation and diagnostics of rolling element bearings with localized faults, J. Vib. Acoust. (2003). [9] P.D. McFadden, J.D. Smith, The vibration monitoring of rolling element bearings by the high-frequency resonance technique: a review, Tribol. Int. 17 (1) (1984) 310. [10] Stephane Mallat, A Wavelet Tour of Signal Processing, second ed., Academic Press, New York, 1999. [11] Donald B. Percival, Andrew T. Walden, Wavelet Methods for Time Series Analysis, Cambridge University Press, England, 2000. [12] Z.K. Peng, F.L. Chu, Application of wavelet transform in machine condition monitoring and fault diagnostics: a review with bibliography, Mech. Syst. Signal Process. 18 (2004) 199221. [13] C.E. Shannon, W. Weaver, The Mathematical Theory of Communication, The University of Illinois Press, Urbana, 1959. [14] S. Blanco, et al., Timefrequency analysis of electroencephalogram series. III. Wavelet packets and information cost function, Phys. Rev. E 57 (1) (1998). [15] L. Zunino, et al., Wavelet entropy of stochastic processes, Physica A 379 (2007) 503512. [16] A. Rosso, et al., Wavelet entropy: a new tool for analysis of short duration brain electrical signals, J. Neurosci. Meth. 105 (2001) 6575. [17] S. Theodoridis, K. Koutroumbas, Pattern Recognition, second ed., Academic Press, New York, 2003.