Professional Documents
Culture Documents
net/publication/257404801
CITATIONS READS
64 72
3 authors:
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by D. H. Pandya on 20 August 2018.
READS
11
3 AUTHORS, INCLUDING:
All in-text references underlined in blue are linked to publications on ResearchGate, Available from: D. H. Pandya
letting you access and read them immediately. Retrieved on: 17 December 2015
Expert Systems with Applications 40 (2013) 4137–4145
a r t i c l e i n f o a b s t r a c t
Keywords: This paper presents a fault diagnosis technique based on acoustic emission (AE) analysis with the Hil-
Acoustic emission bert–Huang Transform (HHT) and data mining tool. HHT analyzes the AE signal using intrinsic mode
KNN functions (IMFs), which are extracted using the process of Empirical Mode Decomposition (EMD). Instead
Statistical feature of time domain approach with Hilbert transform, FFT of IMFs from HHT process are utilized to represent
Ball bearing
the time frequency domain approach for efficient signal response from rolling element bearing. Further,
Hilbert Hung transform
extracted statistical and acoustic features are used to select proper data mining based fault classifier with
or without filter. K-nearest neighbor algorithm is observed to be more efficient classifier with default set-
ting parameters in WEKA. APF-KNN approach, which is based on asymmetric proximity function with
optimize feature selection shows better classification accuracy is used. Experimental evaluation for time
frequency approach is presented for five bearing conditions such as healthy bearing, bearing with outer
race, inner race, ball and combined defect. The experimental results show that the proposed method can
increase reliability for the faults diagnosis of ball bearing.
Ó 2013 Elsevier Ltd. All rights reserved.
0957-4174/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.eswa.2013.01.033
4138 D.H. Pandya et al. / Expert Systems with Applications 40 (2013) 4137–4145
precision and low computational burden as compare to frequency and similarity function were both asymmetric. In fact, they were
domain feature (Zarei, 2012) Significant of multi class relevance directed from a test point to a training point and capture more
vector machine method for fault diagnosis of slow speed rotating information as they realize local level learning about the concerned
machine using acoustic emission technique were investigated training point. A properly chosen value of k can potentially im-
(Yang, Han, & Hwang, 2005). prove the classification results. Thus, in future, we would try to fur-
A new scheme of acoustic based diagnosis (ABD) for rolling ele- ther improve our results by choosing a suitable value of k. We
ment bearing fault diagnosis based on near-field acoustic hologra- would also explore Metric learning and Parzen window based
phy (NAH) and gray level co-occurrence matrix (GLCM) was learning to possibly enhance the performance of the kNN algo-
presented. Proposed modified ABD method with multi class SVM rithm (Bhattacharya, Ghosh, & Chowdhury, 2012).
shown better effectiveness as compare to traditional ABD method In this paper, supervised machine learning methods are per-
(Jiang, Lu, Wu, & Hou, 2012). formed for solving data mining problems for fault classification
Samantha and Al-Balushi have presented a procedure for fault of ball bearing using WEKA. The WEKA default setting of parame-
diagnosis of rolling element bearings through artificial neural net- ters were used for every algorithm to avoid bias and over fitting.
work (ANN). The characteristic features of time-domain vibration Performance of various supervised classifiers such as Bayesian,
signals of the rotating machinery with normal and defective bear- Naïve Bayes, Multi Layer Perceptron, KNN, weighted KNN, J48
ings have been used as inputs to the ANN (Samanta & Al-Balushi, and Random forest are compared to select most suitable classifier.
2003; Widodo et al., 2009). A fault diagnosis scheme based on mul- KNN is selected as best suited classifier. This paper is focused on
ti-class SVMs for rotating machinery were investigated and a com- modified k nearest neighbor classifier for fault diagnosis of ball
parison between two kinds of algorithms, the SVMs and other bearing with low speed using acoustic emission technique and
traditional Neural Networks, such as the self-organizing feature intrinsic mode function. Proposed KNN algorithm is based on in-
map (SOFM), the Learning Vector Quantization (LVQ), the Radial Ba- verse square distance in same class, hence error due to improper
sic Function (RBF) and the Adaptive Resonance Theory (ART) net- selection of K can be compensated. Experimental results show po-
works. The experiments have demonstrated that this approach tential improvement in classification with modified KNN and opti-
can successfully diagnose any condition and the average fault diag- mized feature selection.
nosis accuracy is over 90% (Yang et al., 2005). A new method that
jointly optimizes the feature selection and the SVM parameters
2. Empirical Mode Decomposition
with a modified discrete particle swarm optimization was pre-
sented. This new method can select the best fault features in a
This concept is exploited as utilizing EMD (Empirical Mode
shorter time, improves the performance of the SVM classifier, and
Decomposition) process as a detection filter through IMF (intrinsic
presents fewer errors and a better real-time capacity than the
mode function). EMD computation possesses a band pass filtering
method based on Principal Component Analysis (PCA) and SVM,
nature which is well suited to recover the filtered non linear and
or the method based on Genetic Algorithm (GA) and SVM, as shown
non stationary signals with good detect ability. EMD also optimize
in the application of fault diagnosis of the turbo pump rotor (Yuan &
the signal distortion, which is a limitation of conventional band
Chu, 2007a; Yuan & Chu, 2007b). The application of particle swarm
pass detection filter.
optimization (PSO) combined with other Computational Intelli-
EMD the first step of HHT, is to decompose the complicated
gence (CI) techniques for bearing fault detection in machines was
nonlinear and non-stationary signals into finite number of IMFs.
studied and results were compared with other techniques such as
Intrinsic mode functions (IMFs), which is a kind of complete, adap-
Genetic Algorithm (GA) and Principal Component Analysis (PCA).
tive and almost orthogonal representation for the analyzed signal.
The PSO based approach gave a test classification success rate of
The IMF component is obtained by a sifting process, and it should
98.6–100%, which were comparable with GA and much better than
satisfy two requirements: (1) The number of extrema and the
with PCA. The results show the effectiveness of the selected fea-
number of zero-crossings must be either equal or differ at most
tures and the classifiers in the detection of the machine condition
by one in the whole data set; (2) At any point, the mean value of
(Samanta & Nataraj, 2009). The classification efficiency of features,
the envelope defined by the local maxima and the envelope de-
histogram features seem to be better compared to statistical fea-
fined by the local minima is zeros. Obviously, each IMF only has
tures. Comparing the results of SVM and PSVM where concluded
one frequency value at any moment. Therefore, it is feasible to
that PSVM performs consistently better with Histogram feature. It
compute the meaningful instantaneous frequency once the mul-
has been observed that the researchers are continuously attempt-
ti-component signal is reduced to a collection of mono-component
ing various machine learning methods and combining with differ-
functions through the EMD process. IMF is defined as Eq. (1).
ent feature selection criteria (Sugumaran & Ramachandran, 2011).
Jiang et al. (2012) had performed experiment to evaluate and X
m
xðtÞ ¼ ck ðtÞ þ rm ðtÞ ð1Þ
compare the fuzzy lattice classifier (FLC) with the fuzzy lattice rea-
k¼1
soning (FLR) model as well as some other widely used classification
methods for fault diagnosis of bearing. Experimental results indi- Where, Ck(t) is the kth IMF of the signal x(t) and rm(t) is the final res-
cate that the FLC yields a satisfactory classification performance idue which can be a constant or the signal mean trend. Based on the
with higher computation efficiency than other classifiers. discussion in (Huang et al., 1998). The numerical procedure to ob-
The nearest neighbor is one of the most popular algorithms that tain those IMFs can be concluded with the following steps:
have long been used in pattern recognition, exploratory data anal- (1) Identify the extrema of the data x(t), and form the envelopes
ysis and data mining problems. Some advantages of the kNN algo- defined by the local maxima and minima, respectively, by
rithm are its inherent simplicity, its robustness to noisy training the cubic spline method.
data, especially if the inverse square of weighted distance is used (2) Form the mean values m1(t) by averaging the upper enve-
as the ‘‘distance’’ measure and its effectiveness if the training data lope and lower envelope, and make the differences between
is large. Many researchers have found that the kNN algorithm the data and the mean values to get the first component
achieves good performance in their experiments on different data h1(t) = x(t) m1(t).
sets. Asymmetric proximity functions (distance and similarity (3) If the first component is not an IMF, let h1(t) be the new data
function) in recent years have gained popularity over their tradi- set. Continue the steps (1–2) until the first component is an
tional counterpart. Noted, that affinity based distance function IMF.
D.H. Pandya et al. / Expert Systems with Applications 40 (2013) 4137–4145 4139
(4) The first IMF component is called as c1(t). Let on value of K. that was highlighted by earlier researchers but, as
r1(t) = x(t) c1(t). Continue the steps (1–3) until rn(t) is the new data is closer to class A data this may misguide the classi-
smaller than a predetermined value or becomes a monotonic fication. In general the following steps are performed for this sim-
function that no more IMF can be extracted. ple KNN algorithm (Jiang et al., 2012).
The procedure of the proposed method can be summarized as Chosen of k value: k value is completely up to user. Generally
follows. after some trials a k value which gives the best result is chosen.
Generally, k = 1.
(1) Capturing acoustic emission signals x(t). Distance calculation: any distance measurement can be used for
(2) Performing band pass filtering with Empirical Mode this step. Generally most known distance measurements like
Decomposition. Euclidean and Manhattan distances are used.
(3) Optimize intrinsic mode functions and collect set of IMFs of Distance sort in ascending order: chosen k value is also impor-
definite frequency band. tant at that point. Found distances are sorted in ascending order
(4) Time-frequency domain features extracted through IMF1. and minimum k distance is taken.
Finding k class values: existing classes of k nearest data is
Further, Hilbert transform of first IMF is producing HHT spec- identified.
trum as shown in Fig. 1 this envelop spectrum of healthy and faulty Finding dominant class: in last step identified k classes are
bearing clearing shows the defective bearing condition but with formed a ratio and the class which has maximum ratio is taken.
acoustic signal researchers are not able to identify bearing condi- This taken class gives us the class of desired data.
tion such as vibration signals. In this scenario, significant of ma-
chine learning method for incipient fault diagnosis is increases To eliminate this deficiency, here author would like to present
with acoustic emission signals. modified KNN.
Fig. 1. (a) time response (b) first 4 IMFs (c) FFT of first 4 IMFs and (d) HHT spectrum for healthy and faulty bearing.
test bearing. Its cable is connected to input channel of OROS 3 SER- measurement and default (1 V/g) for absolute measurement. De-
IES analyzer. OROS 3 SERIES is further connected to system. It is fault microphone is selected for type of sensor in analyzer.
analyzed by NVGATE software. Acoustic Sensor is the contacting
type sensor. Four channel OROS analyzer can be used for both 5. Feature extraction and data preparation
acoustic and vibrational analysis. Sampling frequency of 5000 Hz
is selected for analysis. Five seconds data is taken for all readings. Data acquisition of acoustic emission signals are carried out at
Sensitivity of the contact sensor is 100 m V/g for relative low speed range of 250–1500 rpm at interval of 250 rpm. Total 5
D.H. Pandya et al. / Expert Systems with Applications 40 (2013) 4137–4145 4141
P
ak
meanðlÞ ¼ ð7Þ
N
following calculations,
It is generally observed that different features used for classification
peak ¼ maxðak Þ ð6Þ fall in a very wide numerical range. In order to make machine learn-
ing networks perform well, the data must be well processed and
The mean, indicated by l (a lower case Greek mu), is the statisti- properly scaled before inputting to the supervised learning method.
cian’s jargon for the average value of a signal. Add all of the samples The Column data are normalized in the range 0.1–0.9 to minimize
together, and divide by N. It looks like this in mathematical form: the effect of spreadness of attributes.
Some features are sensitive and closely related to the fault, but
others are not. Thus, before the whole feature set is fed into a clas-
sifier, sensitive features providing bearing fault-related informa-
tion must be selected and highlighted and irrelevant features
discarded or weakened to improve the classification performance
and avoid the curse of dimensionality. Each of acoustic emission
signal is proposed to extract nine features. Therefore, a feature
set {Yn,c,j, where ? n = 1, 2, . . . , N, c = 1, 2, . . . , C, j = 1, 2, . . . , J} can be ac-
quired, which is an N-by-C-by-J matrix. Where, N is number of
samples under Cth bearing conditions, C is bearing condition and
Fig. 4. Spall seeded on bearing elements.
J is number of feature. Hence, {Yn,c,j} is expressed as jth feature val-
ued of nth sample under cth bearing condition or class.
In the experiment, 180 data samples were obtained for the five
bearing conditions (HB, ORD, IRD, BD, CD). For each sample, the 9
features are extracted to represent the characteristic information
contained in the sample. Thus, a feature set {Yn,c,j} with
Table 1 36 5 9 feature values is obtained. Proposed optimization con-
Details of rotor-bearing experimental set up. cept is presented in Fig. 5. Features are to arrange in descending or-
der as per their classification efficiency with swapping of feature
Bearing specification SKF 6205
Mass of the rotor (mr) 2.5 kg order if require. Then repeat above outline procedure with differ-
Diameter of inner race with point of 31.12 mm ent number of features (j) and concluded the optimize number of
contact with the rolling element (d) features with maximum classification efficiency. Optimized num-
Diameter of outer race with point of 47 mm
bers of attributes are observed as six out of nine features, which
contact with the rolling element (D)
Radius of each rolling element (qr) 3.965 mm is shown in Fig. 6 with classification accuracy of 96.667%. Classifi-
Radial load (W) 25 N cation accuracy of each feature is tabulated in Table 2.
Outside diameter 52 mm Overall procedure for fault diagnosis is presented in Fig. 7, that
Number of rolling elements (Nb) 9 is described in detail as per following steps.
Angular separation between elements (b) 40°
Spall a seeded defect (l) 300 300 10
Acquire acoustic emission signal from rotor bearing system.
Table 3
Classification accuracy using various supervised classifiers with supervised filters and without filter.
Table 4 has correctly predicted 35, 36, 34, 34 and 34 instances of HB, ORD,
Confusion matrix with weighted KNN classifier. IRD, BD and CD bearing respectively. While modified KNN with
HB ORD IRD BD CD optimized features has correctly predicted 35, 36, 34, 34 and 35 in-
35 1 0 0 0 HB stances of HB, ORD, IRD, BD and CD bearing respectively as shown
0 36 0 0 0 ORD in Table 5. Comparison of weighted KNN and proposed APF-KNN
0 0 34 2 0 IRD with optimized feature is tabulated in Table 6.
0 0 2 34 0 BD Optimal numbers of features (attributes) are selected and mod-
0 0 1 1 34 CD
ified KNN classifier is trained. Now, KNN classifiers are lazy learn-
ers, that is, models are not built explicitly like eager learners (e.g.,
decision trees, SVM, etc.) Thus, building the model is cheap, but
Table 5 classifying unknown objects is relatively expensive since it re-
Confusion matrix with APF-KNN. quires the computation of the k-nearest neighbors of the object
HB ORD IRD BD CD to be labeled. This, in general, requires computing the distance of
35 1 0 0 0 HB the test point to all the training point in data set, which can be
0 36 0 0 0 ORD expensive particularly for large training sets. A number of tech-
0 0 34 2 0 IRD niques have been developed for efficient computation. Here,
0 0 2 34 0 BD author shows the testing concept to reduce the computational time
0 0 1 0 35 CD
without effecting classification accuracy.
Let, test set is consider for each bearing response as healthy
Table 6 bearing and input vector is as x = [F2, F1, F3, F5, F6, F9, 1].
Comparison between weighted KNN and APF-KNN. Test 1: in this test set input vector from healthy bearing is con-
sider, hence corresponding confusion matrix as shown in Fig. 8(a).
Weighted KNN APF-KNN attribute
Which is clearly shows the classification accuracy of 100% with
Correctly classified instances 173 (96.1111%) 174 (96.6667%)
bearing condition HB.
Incorrectly classified instances 7 (3.8889%) 6 (3.3333%)
Total number of instances 180 180
Test 2: in this set input vector from bearing with combined de- and estimation of defect size. Mechanical Systems and Signal Processing, 20,
1537–1571.
fect is consider, hence corresponding confusion matrix clearly
Bhattacharya, G., Ghosh, K., & Chowdhury, A. S. (2012). An affinity-based new local
shows unclassified data and as shown in Fig. 8(b) indicate that test distance function and similarity measure for kNN algorithm. Pattern Recognition
data classify at first column and fifth raw rather than at fifth col- Letters, 33, 356–363.
umn and fifth raw. Hence user can simply classified that the bear- He, Y., Zhang, X., & Friswell, M. I. (2009). Defect diagnosis for rolling element
bearings using acoustic emission. Journal of Vibration and Acoustics, 131.
ing of 5th class inserted as 1st class. 061012-1.
Huang, N., Shen, Z., Long, S., Wu, M., Shih, H., Zheng, Q., et al. (1998). The empirical
mode decomposition and the Hilbert spectrum for nonlinear and non-
7. Conclusion stationary time series analysis. Proceedings of the Royal Society of London Series
A, 454, 903–995.
In this paper, fault diagnosis of ball bearing is reported using Jiang, W., Lu, W., Wu, H., & Hou, J. (2012). A fault diagnosis scheme of rolling
element bearing based on near-field acoustic holography and gray level co-
acoustic signal and intrinsic mode function. These intrinsic fea-
occurrence matrix. Journal of Sound and Vibration, 331, 3663–3674.
tures are inserted to seven supervised learning methods to com- Jiang, Shengyi., Pang, Guansong., Meiling, Wu., & Kuang, Limin. (2012). An improved
pare and selected KNN as the most efficient supervised learning K-nearest-neighbor algorithm for text categorization. Expert Systems with
Applications, 39, 1503–1509.
method with three different filters as well. Then to further improve
Lei, Yaguo., & Zuo, Ming. J. (2009). Gear crack level identification based on weighted
the classification accuracy, modified KNN algorithm is proposed K nearest neighbor classification algorithm. Mechanical Systems and Signal
based on asymmetric proximity function (APF) in simple way, as Processing, 23, 1535–1547.
APF have gained recognition in recent years (McFee B., 2010). Here, McFee B., & Lanckriet G. (2010). Metric learning to rank. In Proceedings of the ICML
(pp. 775–782). Haifa, Israel.
information is directed from test point to training data set and Mehmet Cigdem, I., & Mutlu, A. (2010). A hybrid classification method of k nearest
hence more realistic information at local level acquire with APF. neighbor, Bayesian methods and genetic algorithm. Expert Systems with
Appropriate value of K plays a significant roll in performance of Applications, 37, 5061–5067.
Mitra, P., Murthy, C. A., & Pal, S. K. (2002). Density-based multiscale data
KNN as concluded by earlier researchers (Mitra et al., 2002), in this condensation. IEEE Transactions on Pattern Analysis and Machine Intelligence,
pffiffiffi
paper modified KNN consider k ¼ n, where n is the number of 24(6), 734–747.
data used for training purpose for each class or bearing condition Morhain, A., & Mba, D. (2003). Bearing defect diagnosis and acoustic emission.
Journal of Engineering Tribology Part I: Mechanical Engineering, 217(4), 257–
and value of k will be nearest odd number. 272.
Experimental results clearly shows that proposed APF-KNN Samanta, B., & Al-Balushi, K. R. (2003). Artificial neural network based fault
algorithm with optimize feature has outperform the KNN algo- diagnostics of rolling element bearings using time-domain features. Mechanical
Systems and Signal Processing, 17(2), 317–328.
rithm with 96.6667% without compromising classification time
Samanta, B., & Nataraj, C. (2009). Use of particle swarm optimization for machinery
and computational memory requirement. KNN is instant based fault detection. Engineering Applications of Artificial Intelligence, 22(2), 308–
algorithm, which theoretically require more classification time 316.
Sugumaran, V., & Ramachandran, K. I. (2011). Effect of number of features on
and computational memory but both the draw backs are almost
classification of roller bearing faults using SVM and PSVM. Expert Systems with
negligible in our scenario. State of art classification can be achieved Applications, 38, 4088–4096.
by storing 36 samples with 9 feature vectors in one class and that Tandon, N., & Chaudhary, A. (1999). A review of vibration and acoustic
no significant improvement can be expected with optimized 6 fea- measurement methods for the detection of defects in rolling element
bearings. Tribology International, 32, 469–480.
ture vectors. Widodo, Achmad, Kim, Eric. Y., Son, Jong-Duk, Yang, Bo-Suk, Tan, Andy. C. C., Dong-
The results of investigation revalidate that HHT based IMF com- Sik, Gu., et al. (2009). Fault diagnosis of low speed bearing based on relevance
ponent of acoustic signal from different bearing conditions can be vector machine and support vector machine. Expert Systems with Applications,
36, 7252–7261.
effectively applied to diagnose the incipient fault conditions of Yang, B. S., Han, T., & Hwang, W. W. (2005). Fault diagnosis of rotating machinery
bearing. based on multi-class support vector machines. Journal of Mechanical Science and
Technology, 19(3), 846–859.
Yoshioka, T., & Fujiwara, T. (1984). Application of acoustic emission o detection of
Acknowledgment rolling bearing failure. ASME Production Engineering Division Publication, 14,
55–76.
This work is financially supported by the Ministry of Human Re- Yuan, S. F., & Chu, F. L. (2007a). Fault diagnosis based on support vector machines
with parameter optimization by artificial immunization algorithm. Mechanical
source Development (MHRD), Government of India.
Systems and Signal Processing, 21(3), 1318–1330.
Yuan, S. F., & Chu, F. L. (2007b). Fault diagnosis based on particle optimization and
References support vector machines. Mechanical Systems and Signal Processing, 21(4),
1787–1798.
Abdullah, M., Al-Ghamd & Mba, D. (2006). A comparative experimental study on the Zarei, J. (2012). Induction motors bearing fault detection using pattern recognition
use of acoustic emission and vibration analysis for bearing defect identification techniques. Expert Systems with Applications, 39, 68–73.