Wavelet in TCM Review

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/223524626
Wavelet analysis of sensor signals for tool condition monitoring: A review and
some new results
Article in International Journal of Machine Tools and Manufacture · June 2009

DOI: 10.1016/j.ijmachtools.2009.02.003
CITATIONS READS
372 1,151
3 authors, including:
Kunpeng Zhu Geok-Soon Hong

Hefei Institute of Physical Sciences, Chinese Academy of Sciences National University of Singapore
89 PUBLICATIONS 2,656 CITATIONS 164 PUBLICATIONS 5,391 CITATIONS
SEE PROFILE SEE PROFILE
All content following this page was uploaded by Kunpeng Zhu on 17 April 2020.
The user has requested enhancement of the downloaded file.

Author's personal copy
ARTICLE IN PRESS
International Journal of Machine Tools & Manufacture 49 (2009) 537–553
Contents lists available at ScienceDirect
International Journal of Machine Tools & Manufacture

journal homepage: www.elsevier.com/locate/ijmactool
Wavelet analysis of sensor signals for tool condition monitoring:

A review and some new results
Zhu Kunpeng, Wong Yoke San, Hong Geok Soon
Department of Mechanical Engineering, National University of Singapore, Singapore 119260, Singapore
a r t i c l e in fo abstract
Article history: This paper reviews the state-of-the-art of wavelet analysis for tool condition monitoring (TCM).
Received 9 May 2008 Wavelet analysis has been the most important non-stationary signal processing tool today, and
Received in revised form popular in machining sensor signal analysis. Based on the nature of monitored signals, wavelet
10 February 2009
approaches are introduced and the superiorities of wavelet analysis to Fourier methods
Accepted 12 February 2009
Available online 27 February 2009
are discussed for TCM. According to the multiresolution, sparsity and localization properties of wavelet
transform, literatures are reviewed in five categories in TCM: time–frequency analysis of machining
Keywords: signal, signal denoising, feature extraction, singularity analysis for tool state estimation, and
Wavelet density estimation for tool wear classification. This review provides a comprehensive survey of the
Tool condition monitoring
current work on wavelet approaches to TCM and also proposes two new prospects for future studies in
this area.
& 2009 Elsevier Ltd. All rights reserved.
Contents
1. Overview of tool condition monitoring. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537

1.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
1.2. TCM as a pattern recognition problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 538
2. Wavelet, wavelet transform, and properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 539
2.1. Limitation of time–frequency resolutions of Fourier methods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 539
2.2. Wavelet and continuous wavelet analysis (CWT). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 540
2.3. Discrete wavelet transform (DWT) and wavelet packet decomposition (WPD) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
2.4. Useful properties of wavelet transform for TCM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
3. Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
3.1. Time–frequency analysis of TCM sensor signals with wavelet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 542
3.2. Wavelet denoising in TCM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 544
3.3. Feature extraction and dimension reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 545
3.4. Singularity analysis for tool wear detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 547
3.5. Wavelet probability density estimation for tool
state classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 549
4. Conclusion and future studies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 550
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 550
1. Overview of tool condition monitoring tool wear or abruptly as tool fracture or breakage [1]. In
tool condition monitoring, the aim is to apply appropriate
1.1. Introduction sensor signal processing and pattern recognition techniques to
identify and predict the cutting tool state, so as to reduce
During machining, the contact between the cutting tool, loss brought about by tool wear or tool failure. An effective
workpiece, and the chips imposes pressure on the tool and tool condition monitoring (TCM) system can improve productivity
causes the shape of the tool to change, either gradually as and ensure workpiece quality, and hence, has a major influence
on machining efficiency [2]. Tool condition monitoring has
been extensively studied by many researchers since the late
Corresponding author. 1980s. Many of the reported research works are reviewed in
E-mail address: mpezhuk@nus.edu.sg (K.P. Zhu). [3–5].
0890-6955/$ - see front matter & 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijmachtools.2009.02.003
ARTICLE IN PRESS
538 K.P. Zhu et al. / International Journal of Machine Tools & Manufacture 49 (2009) 537–553
Nomenclature WTMM wavelet transform modulus maxima

WVD Wigner–Ville distribution
AE acoustic emission
AR autoregressive Symbols
ART adaptive resonance theory
CQF conjugate quadratic filters f(t) any mathematical signal f(t)ALR2
CWD Choi–Williams distribution x(t) any sensory signal from machining
CWT continuous wavelet transform y(t) extracted features
DFT discrete Fourier transform f^ ðoÞ Fourier transform of f(t)
DWT discrete wavelet transform Ts sampling interval
EMD empirical mode decomposition fs sampling frequency
FDR Fisher’s discriminant ratio Ci class i
FFT fast Fourier transform /x(t),y(t)S inner product of signal x(t) and y(t)
FWT fast wavelet transform u position parameter of wavelet function
GMM Gaussian mixture models s scale parameter of wavelet function
HMM hidden Markov models st the resolution of time
ICA independent component analysis so the resolution of frequency
KLT Karhunen–Loeve transform c(t) wavelet function
LE Lipschitz exponent j(t) scaling function
LDA linear discriminant analysis h(n) low-pass filter
MLP multilayer perceptron g(n) high-pass filter
MRA multiresolution cj,k scaling coefficient
NN neural networks dj,k wavelet coefficient
PCA principal component analysis Wj,n,k wavelet packet
pdf probability density function n
dj;k wavelet packet coefficient
PSD power spectrum density Q quality factor
SNR signal-to-noise ratio a Lipschitz exponent
SOM self-organizing map Si the covariance matrix of class i
STFT short-time Fourier transform SW within-class covariance matrix
TCM tool condition monitoring SB between-class covariance matrix
WPD wavelet packet decomposition
WT wavelet transform
Since tool condition is typically defined according to the and motor/feed current [27–32] have been the most employed
geometrical changes in the tool, direct monitoring methods such and reported for TCM. Detailed works on the design and
as vision and optical approaches, which measure the geometric implementation of these indirect approaches for TCM have been
parameters of the cutting tool, have been developed [6–8]. The reported in [33–35].
direct methods have advantages of capturing actual geometric
changes arising from wear of tool. However, direct measurements
are very difficult to implement because of the continuous contact 1.2. TCM as a pattern recognition problem
between the tool and the workpiece, and almost impossible due to
the presence of coolant fluids. The difficulties severely limit the The problem of TCM can be considered as a typical
application of direct approach. The indirect approaches are pattern recognition problem. The objectives of TCM
achieved by correlating or deducing suitable sensor signals to can be formally specified to be a search for the most probable
tool wear states. They have the advantages of less complicated state Ci given the extracted measurable signal feature y(t)
setup and suitability for practical application. This paper focuses at time t. This is a dynamic inference problem since the
on indirect approaches. For indirect approaches, tool condition is tool state is not estimated only with prior knowledge, but also
not captured directly, but estimated from the measurable signal adapt to the current features. This is somewhat of Bayesian
feature. This signal feature is extracted through signal processing inference [36].
steps (Fig. 1) for sensitive and robust representation of its Hence, as the pattern recognition problem, the aim of TCM is to
corresponding state. find,
Indirect methods such as those based on sensing of the cutting TCM : arg max pðC i jyÞ (1)
forces [9–15], vibrations [16–20], acoustic emission (AE) [21–26], i
Machining Signals Signal Processing Classifier Tool State
Setup Sensors Force Amplitude analysis features Time series estimation Wear
AE Fourier analysis LDA Chipping
Vibration Wavelet analysis NN Breakage
Current Statistical moments Clustering Failure
Image Time series HMM Chatter
Fig. 1. The framework of TCM.

ARTICLE IN PRESS
K.P. Zhu et al. / International Journal of Machine Tools & Manufacture 49 (2009) 537–553 539
or in the physical form Markov models (HMMs) [50,51], support vector machine (SVM)
[52], and rough set [53] have also been studied and applied to
TCM : arg max pðtool statejsignal featuresÞ (2) TCM by many researchers.
tool state i
Table 1 lists them according to their classification approaches.
The TCM system can be achieved in a three-step procedure as Note that the list is not totally exhaustive but serves to be
pattern recognition (Fig. 2). The sensor signal x(t) is firstly pre- representative of known TCM approaches, where most of them
processed to remove the noise and prepare the data for feature also involve wavelet applications to one of the three stages of TCM
extraction. Then the information relevant to pattern classification (Fig. 2).
is extracted from x(t) to a feature vector y(t). The task of feature
extraction is to enhance the characteristics of the various tool
wear classes and suppress or filter off the normal background. The 2. Wavelet, wavelet transform, and properties
final stage is state classification. Feature vector y(t) is assigned to
one of the K tool wear state, C1,C2,y,CK, by the classifier based on 2.1. Limitation of time–frequency resolutions of Fourier methods
a certain type of classification criteria.
Earlier study on TCM classifier is mainly carried out with time- We call any square integrable real function f(t)AL2(R) a signal.
series analysis [9,37,38]. With these methods, a threshold value For the signal f(t), the Fourier transform f^ ðoÞ is obtained by the
needs to be set between the normal and abnormal tool states. inner product of f(t) with a sinusoidal wave ejot,
However, the threshold value varies with cutting conditions and is D E Z 1
difficult to determine. To improve the performance of TCM, more f^ ðoÞ ¼ f ðtÞ; ejot ¼ f ðtÞejot dt (3)
advanced methods have been developed. Neural networks (NNs) 1
are most studied and gained most success in practical applications It transforms the signal f(t) from the time domain to the
[39–44], due to its capability in learning and non-linear mapping frequency domain o and is viewed as the basis of modern signal
of features and tool state. Besides NNs, other pattern recognition processing. The fast Fourier transform (FFT) [54] is the standard
methods, such as fuzzy clustering approaches [45,46], linear method for observing signals in the frequency domain and has
discriminant analysis (LDA) [37], Gaussian mixture models been widely studied in TCM, such as those [11,55,56]. In spite of its
(GMMs) [47], combination of regression and neuro-fuzzy techni- earlier popularity, Fourier transform has certain serious theore-
ques [48], self-organizing feature maps (SOM) [8,49], hidden tical drawbacks in processing machining signals. This is because
the f^ ðoÞ is the integration of f(t) for all times tA(N,+N)
(function 3), and this globally inclusive of information makes it
difficult to analyze any local property of f(t) from f^ ðoÞ. To
overcome this limitation, Gabor [57] introduced a sliding window
Fig. 2. TCM as a pattern recognition system. function g(t) to the Fourier transform and obtains a localized
Table 1
Tool condition estimation scopes.
Approaches References Comments
Time series AR, ARMA [9,37,38,127] Linear, physical meaning, good for stationary machining like
turning, not good for non-stationary machining like milling, need
set threshold for classification
Neural networks MLP [12–14,19,41–44,120,126,130] Iterative MSE optimization, sensitive to network structure, non-
linear classification, slow training
SOM [8,40,49,128] Non-linear and iterative clustering, suitable for low-dimension
feature space
ART [24,40,67,68,83,131] Based on competitive learning: fast incremental learning ability,
good self-adaptive ability
SVM [52] Maximizing the margin between classes with minimum number
of support vectors, metric dependent, non-linear, good
generalization, slow training
Others [122,123] Sensitive to training parameters, non-linear classification, robust
to outliers
Pattern recognition k-means [12,13] K-clusters, the nearest mean decides the cluster, good for
Gaussian signal with equal covariance
Fuzzy methods [12,13,45,46,76,132] Need initializing clusters and class membership
Gaussian mixture models [47,148] Each state is assumed to be drawn number of underlying
Gaussian distributions, soft membership, better than k-means
clustering, need estimate components
PCA/KLT [91,92,96,126] Linear, second-order statistics based on eigenvector
decomposition, good for Gaussian signal
LDA [12,13,23,37,38,61,96,124] Supervised linear classifier, using MSE for optimization, better
than PCA for classification, need Gaussian assumption of signal
Stochastic models Hidden Markov models [50,51,129] Simple structure, good in generalization, good in non-linear and
non-stationary machining signals, need to train many small
models
ARTICLE IN PRESS
time–frequency atom f: tooth signal, but instead it spread the information across the
jo t
transformed signal. Wavelet analysis overcomes the drawbacks of
fu;o ðtÞ ¼ e gðt uÞ (4) Fourier methods and permits adaptive time–frequency represen-
The resultant transform is named short-time Fourier transform tation. Gong et al. [62] have shown that the wavelet analysis is
(STFT): more sensitive and reliable than the Fourier analysis for
Z 1 recognizing the tool wear states in turning. Yoon and Chin [63]
also verified the reliability of the wavelet transform method
ðSTFTÞðf ðtÞÞ ¼ f ðtÞ; fu;o ¼ f ðtÞgðt uÞejot dt (5)
1 compared the spectra method of FFT. The signal processing
The corresponding energy density |STFT(f(t))|2is called a approaches that deal with non-stationary signals are more
spectrogram, which is widely used for time–frequency analysis appropriate for process monitoring.
before wavelet, and applied to TCM [49,58].
However, the frequency resolution s2o and time resolution s2t 2.2. Wavelet and continuous wavelet analysis (CWT)
are constant for both time and frequency (Fig. 3(c)), and according
to Heisenberg uncertainty principle [59], To meet the needs for adaptive time–frequency analysis in
2 2 1 applied mathematics, physics, and engineering, the wavelet
s s
t o X4 (6)
theory was developed in the late 1980s by Mallat [64,59], and
equality holds if f(t) is a Gaussian. It states that high resolution Daubechies [65,66]. It has been widely used to analyze machining
both in frequency and time cannot be attained at the same time. signals for tool wear monitoring since Tansel et al. [67,68].
As shown in Fig. 3, while the original time signal f(t) has infinite Let cs,u(t), s,uAR,s40, be a family of functions defined as
time resolution (Fig. 3(a)) but no frequency information, the translations and re-scales of a single function cs;u ðtÞ 2 L2 ðRÞ [59],
Fourier transform f^ ðoÞ has infinite frequency resolution (Fig. 3(b))
1 tu
but provides no time information. The time–frequency resolution cs;u ðtÞ ¼ pffiffi c (7)
s s
of STFT is constant (Fig. 3(c)).
On the other hand, the above-mentioned approaches assume where s is the scaling parameter and u the position parameter.
that the sensor signals are stationary. However, due to the nature The wavelet cs,u(t) has the following basic properties:
of manufacturing processes, the signals are usually non-stationary Z 1 Z 1
[12,60]. Constant time and frequency resolutions of STFT are not cðtÞ dt ¼ 0; c2 ðtÞ dt ¼ 1 (8)
1 1
suitable for the analysis of non-stationary signal. For example,
Mori et al. [61] took the FFT of the thrust force signal and found These properties indicate that the wavelet is a small wave:
that the spectra cannot capture the localized aspect of the saw- oscillate around zero (zero mean) and has limited support area
(finite energy), as shown in Fig. 4(a). The wavelet cs,u(t) has to
meet the admissibility condition for the transformation to be
invertible [65].
ω Continuous wavelet transform of signal f(t) is defined as
Z 1
ω 1 tu
σt W c f ðs; uÞ ¼ f ðtÞ; cs;u ðtÞ ¼ pffiffi f ðtÞc dt (9)
s 1 s
The time–frequency resolution of wavelet transform is illu-
ω2 σω
strated in Fig. 4(b). The signal is localized in the area with time
t width Dt : ½u0 ð1=2Þa0 st ; u0 þ ð1=2Þa0 st and frequency width
σt
Do : ½ðZ=a0 Þ ðso =2a0 Þ; ðZ=ZÞ þ ðso =2a0 Þ. Compared with the
ω
ω1 σω STFT, whose time–frequency resolution is constant Fig. 3(c), the
time–frequency resolution of the wavelet transform (WT) de-
pends on the frequency of the signal. At high frequencies, the
wavelet reaches at a high time resolution but at a low-frequency
resolution; whereas at low frequencies, high-frequency resolution
0 u1 u2 t and low time resolution can be obtained.
t
On the other side, the quality factor of WT is kept constant
Fig. 3. Time–frequency resolutions of Fourier transform and wavelet transform. in the time–frequency plane. For mother c(t/a), the quality
ω aσ t
Morlet Wavelet Daubechies Wavelet
1 2 η σω
a a
0.5 1
ψ (u,s)
ψ (u,s)
a0σ t
0 0
η σω
-0.5 -1 a0 a0
-1 -2
-50 0 50 0 2 4 0 u u0 t
u u
Fig. 4. Typical wavelets and adaptive time–frequency resolution of WT: u is position of the wavelet and s the scale of the wavelet.
ARTICLE IN PRESS
factor defined, decomposition [69] is a generalization of wavelet decomposition

at higher frequencies. In the wavelet packet decomposition, each
Q a ¼ Central frequency=bandwidth ¼ ðZ=aÞ=ðso =aÞ ¼ Z=so
approximate and detail coefficients are recursively decomposed.
and for c(t/a0), The library of wavelet packet basis functions fW n g1
n¼0 can be
Q a0 ¼ Central frequency=bandwidth ¼ ðZ=a0 Þ=ðso =a0 Þ ¼ Q a obtained from a given W0 as follows:

pffiffiffi X
As a result, no matter what u is (u40), c(t/a), and c(t/a0) W j2n ¼ 2 hðkÞW n ð2t kÞ (15)
are constant with quality factor. Fig. 4(d) illustrates the k
bandwidth and central frequency varies with the different scale pffiffiffi X
u. Such adaptability of time–frequency analysis is called ‘‘math- W j2nþ1 ¼ 2 gðkÞW n ð2t kÞ (16)
ematical micro-scope’’ and benefits the signal analysis crossing k
scales. This multiscale analysis is preferred in TCM to get both where the functions W0 and W1 are set to the scaling function f(x)
coarse global information and fine localized details when and the mother wavelet function c(x), respectively. The imple-
necessary. mentation of the wavelet packets leads to a tree-structured
In physical interpretation, the coefficients of the wavelet decomposition, thereby implying that both the outputs of the
transform indicate the variation of energy of the signal with time low-pass and high-pass filters are recursively decomposed.
and frequency. In engineering applications, the square of the The wavelet packet coefficients are then produced from the
coefficients of the CWT is often called as scalogram, defined integral:
as Eq. (10), which has been used for machinery fault diagnostics, Z
and TCM. j
dn;k ¼ W j;n;k ðtÞf ðtÞ dt (17)
2
SC t ðs; u; cÞ ¼ W c f ðs; uÞ (10)
It should be emphasized that Eq. (18) allows many possible
combinations of wavelet packet functions to be selected in order
to optimally characterize the signal. Several criteria have been
2.3. Discrete wavelet transform (DWT) and wavelet packet developed to select the best basis for these purposes [59,70].
decomposition (WPD) Due to the benefits of wavelet decomposition, wavelet
methods have been studied in all aspects in TCM, such as
Continuous wavelet transforms are recognized as effective time–frequency analysis, signal denoising, feature extraction and
tools for both stationary and non-stationary signals. However, it compression, or directly used as classifier for TCM. Basic theories
involves much redundant information and is computationally of these approaches are introduced and literatures are reviewed
very slow. Discrete wavelet transform was developed by Mallat in Section 3.
with fast algorithm based on the conjugate quadratic filters (CQF)
[64]. Wavelet and scaling functions at different scales are
generated from a single scaling function f(t) with two-scale 2.4. Useful properties of wavelet transform for TCM
difference equations [59]:
pffiffiffi X The most important property of wavelet useful in tool
fðtÞ ¼ 2 hðkÞfð2t kÞ (11) condition monitoring is its sparse representation of signal. The
k
wavelet expansion coefficients cj,k and dj,k decay rapidly with
pffiffiffi X increase in j and k, and only a few large coefficients exist
cðtÞ ¼ 2 gðkÞfð2t kÞ (12)
while the others are small. By setting a suitable threshold, the
k
undesired noise is filtered. This is the essence of wavelet
where g(k) ¼ (1)kh(1k), and the h(k) and g(k) are viewed as denoising, and compression [71,72]. Another important property
filter coefficients of low-pass and high-pass filters, and l is the of wavelet transform that determines its applications is its
filter length. f(t) And c(t) are scaling and wavelet functions at localization property, as discussed in Section 2.1. Unlike Fourier
scale j ¼ 1, respectively. In an orthogonal wavelet expansion, a set transform that spans the entire time period, wavelet transform
of recursive relationships governs scaling and wavelet coefficients localize the time and frequency description of the signal, and
at different scales and translations as follows: reveals the signal behavior in certain time and its corresponding
X
cj;k ¼ hl2k cjþ1;l (13) frequency property, which is generally useful for uncovering
l different localized features associated with various different tool
X states.
dj;k ¼ g l2k cjþ1;l (14)
l
where cj,k, dj,k are scaling and wavelet coefficients derived from 3. Applications
the projection of the signal onto the space of scaling fj,k(t) and
wavelet functions cj,k(t), respectively. We first presents a review of the time–frequency analysis of
Fig. 5(a) illustrates the 5-level MRA analysis of a cutting force machining signals, signal denoising, and feature extraction of
signal sampled at 6000 Hz. With a 5-level MRA analysis, the wavelet applications for TCM. These three applications are
corresponding frequency bands are separated as illustrated in studied in most of the literature reviewed. Two new approaches,
Fig. 5(b). By DWT, the signal f(t) is decomposes into two parts: called singular detection and density estimation with wavelet
low-pass approximation coefficients and high-pass detail coeffi- for TCM, are also introduced with real experiments and
cients. The next step then decomposes the new approximation discussed for TCM. Though few papers are found on these two
coefficients. approaches on TCM, papers related to machinery condition
The DWT lead to a loss of useful information at high frequency monitoring and fault diagnosis are reviewed to show the
because successive details are no longer analyzed. We need to prospect of these applications in TCM. Table 2 lists the references
double the sampling rate for higher frequency analysis, which reviewed according to different wavelet features used for TCM in
however involves more data and computation. Wavelet packet this paper.
ARTICLE IN PRESS
Cutting force and it's multiresolution decomposition

10
5
0
-5
0 200 400 600 800 1000 1200 1400 1600 1800
Approximation level 1 Detail level 1
10 5
5
0
0
-5 -5
500 1000 1500 500 1000 1500
10 5
5
0
0
-5 -5
500 1000 1500 500 1000 1500
10 5
5
0 0
-5 -5
500 1000 1500 500 1000 1500
10 5
5 0
0
-5 -5
500 1000 1500 500 1000 1500
10 5
5
0
0
-5 -5
500 1000 1500 500 1000 1500
0 47Hz 94Hz 188Hz 375Hz 750Hz 1500Hz 3000Hz
V5 W5 W4 W3 W2 W1
V4
V3
V2
V1
V0
Fig. 5. (a) Five-level MRA analysis of cutting force: abscissa is the location of sampled cutting force, and ordinate is the force amplitude in Newton. (b) Frequency band
separation of 5-level MRA analysis.
3.1. Time–frequency analysis of TCM sensor signals with wavelet localized features for TCM. Minimizing the effect of overlapping
and improving the localization are still problems in CWT. These
In the time–frequency analysis of TCM sensor signals, the are partially studied in [75] with an exact WT for gear fault
scalogram SCt(a,b;c), wavelet coefficient dj,k, scaling coefficient detection.
cj,k, wavelet packet coefficients djk, and their wavelet domain DWT is preferable in the time–frequency analysis because of
statistics (i.e.: mean, variance, etc.) are used as criterion in no redundancy and fast computation. Gong et al. [62] applied
condition discrimination in TCM after some manipulation. The DWT to monitor the flank wear states in turning. It was found that
corresponding literatures reviewed in this section are summar- the 5-level coefficients were sensitive to the flank wear and
ized in Table 3. cutting conditions. The normalized 5-level mean wavelet coeffi-
Yesilyurt [73] uses the scalogram and mean frequency of cients were used as parameters of the flank wear state recogni-
scalogram of vibration signals in milling breakage detection. tion. The experimental results showed that it was more reliable
The mean frequency of scalogram characterized the energy than FFT analysis in turning TCM. Li et al. [76] discussed a tool
density of the signal in a certain period. It was found that the breakage monitoring system based on DWT of acoustic emission
feed rate was highly correlated to the mean frequency of and feed current signals. The experimental results show overall
scalogram, and the mean frequency variation is quite sensitive 98.5% reliability and the good real-time monitoring capability of
to the presence of fault. Khraisheh et al. [74] found that CWT is the DWT for detecting tool breakage during drilling. Fu et al. [77]
suitable for analyzing the transient in primary chatter. The identified saw tooth and screeching behavior in the thrust force
transient boundary and the built-up edge were successfully signal by using convolution masks to extract various features from
identified by wavelet transform. With CWT, as discussed earlier, the DWT coefficients. The drawback of this method is that the
it is a redundant transform. We generally encounter the problem profiles of the convolution masks are derived from idealized
of overlapping, as a large amount of redundant information exists simulated signals, and the shapes of the simulated signals
after CWT; the overlapping may have the effect of smearing the significantly influence the output results. Suh et al. [78] developed
ARTICLE IN PRESS
Table 2
Overview of the selected references with wavelet applications.
Wavelet Wavelet features Tool condition Monitored signal Reference
CWT Scalegram: SCt(a,b,c) mean, variance of Scalegram Drill breakage Current [29]
Drill wear Current [29]
Mill breakage Vibration [73]
Milling chatter Vibration [74]
DWT cj,k, dj,k, Mean, variance moments, and cumulants of DWT Drill breakage Current [29]
Force [61,67,68,77]
AE [76,132]
Drill wear Force [47,81]
Vibration [130]
AE [132]
Curremt [29]
Milling chatter Force [63,78,79]
Milling tool wear Force [80,125]
Milling fracture Current [111]
Turning tool wear AE [122,123]
Force [62,124,126]
Vibration [120,129]
Turning failure Force [110]
Grinding too wear Force [109]
n
WPD dj;k , Mean, variance moments, and cumulants of WPD Drilling wear Vibration [60]
Drilling breakage Current [46]
Turning chatter Force [60]
Turning chipping/breakage AE [24,82]
Turning wear AE [24,82]
Current [121]
Vibration [128]
Force [128,131]
MP Pm1 D E
Drilling failure Force [85]
f ¼ n¼0 Rn ; fgn fgn þ Rm
Table 3
Time–frequency properties of the transforms.
Wavelet Reference Transform Suitability Comments
CWT [73–75] Linear Stationary, non-stationary Redundant transform, computationally slow

DWT [62,76–81] Linear Stationary, non-stationary No redundancy, computationally fast, only low-
frequency components iteratively decomposed
WPD [24,60,82,83,121,127,131] Linear Stationary, non-stationary Redundant transform, both low- and high-frequency
components iteratively decomposed
MP [85] Linear Stationary, non-stationary No need to be orthogonal basis, acted also as a classifier
STFT [49,58,62,63] Linear Stationary Can not adaptively change time and frequency resolution
WVD/CWD [86–89] Quadratic Stationary, non-stationary Accurate time–frequency distribution, but involves
cross-interference terms
KLT [91,92] Linear Stationary A PCA approach, can not capture signal’s non-Gaussian
property
a DWT-based algorithm to capture the variations of periodicity, [80], which claimed that the variance of wavelet coefficients are
scaling coefficients, and wavelet coefficients. The three para- sensitive to tool wear and little influenced by the variation of
meters were used to characterize stable, conditionally stable, and working conditions, and thus provide a robust description of tool
unstable milling conditions. It was claimed that this approach can wear in milling. It is reasonable to use the deviation as standard to
accurately detect the transition of the system from one state to detect chatter because the signal amplitude varies largely before
the other. and after the chatter period, but how to generate a robust decision
Yoon and Chin [63] applied with the standard deviation ratio of measure (i.e. threshold) is still challenging. Luis Alfonso et al. [81]
DWT coefficients of cutting force to detect chatter. It was found described a method using current signals to estimate the tool
that the detail coefficient parameters of the third or fourth level conditions by using the DWT of cutting force, and through an
were desirable for detection of chatter with spindle speed of autocorrelation algorithm to evaluate the tool wear in the form of
500–1300 rpm in end milling. The identification of chatter an asymmetry weighting function. The authors reported that the
through wavelet was also investigated by Berger et al. [79] who asymmetry value increases according to the states of the tool
analyzed the ratios of the mean absolute deviations of detail wear. This approach is promising since it uses the current signals,
coefficients of cutting forces. Similar study was also reported in which is easy to implement and do not interrupt the machining.
ARTICLE IN PRESS
−
⎯⎯⎯
→ ⎯⎯⎯⎯⎯⎯
→ ⎯⎯⎯→
Fig. 6. Wavelet denoising scheme.
The problem is that the current signal may not be sensitive in all
conditions.
Wavelet transform can be very helpful if it is used as a signal
separation tool in TCM. Niu et al. [24] and Niu [82] separate the
acoustic emission signal into burst and continuous components
by WPD. The burst signal is suitable for the detection of transient
tool conditions, such as chipping and tool fracture, etc., and the
continuous signal is suitable for determination of tool wear. This
approach is much alike the one presented in [83]. This is
promising as only one signal is used to carry different monitoring
tasks.
The above WT approaches generally need feature extractors Fig. 7. (a) Hard thresholding. (b) Soft thresholding. Wavelet denoising.
followed by classifiers for TCM. A modified WPD analysis, the
matching pursuit (MP) [84], handles both of the problems in an
integrated manner. It decomposes the signal into a linear spectra has no discriminatory power at all. But denoising is
expansion of waveforms that are selected from a dictionary of generally not concerned in TCM area. This may due to the high
basis functions. It chooses waveforms that best match the signal signal to noise ratio (SNR) in conventional machining, and the
structure iteratively. As a pattern recognition method, MP adapts noise imposes little effect on the final decision. On the other hand,
to specific machining monitoring tasks and do not need feature in statistical analysis such as mean or moving average [97], the
extraction or classification. Fu et al. [85] applied the MP to predict noise is diminished by the averaging. But in high precision and
small drill bit failure, with different types of drilling behavior from micro machining, the machining signal is typically very small, and
the thrust force. It was revealed that the MP approach performed as a result the SNR is relatively low [93,95,98]. The noise has to be
satisfactorily with Gaussian, Haar, and Gabor wavelet dictionaries removed before further analysis for TCM.
in detecting small drill bit pre-failure. This approach simplifies the The model of signal with noise is [99]
structure of TCM. The problem is that the threshold selection
yðt i Þ ¼ f ðt i Þ þ ðt i Þ (18)
scheme in the wear detector is hard to generalize.
In some cases, the energy variation of non-stationary signal is where the function f(t) represents the desired signal, while the
examined for TCM. Quadratic time–frequency methods such as remaining part ei is the noise. For the wavelet denoising, applying
the Wigner–Ville distribution (WVD) and the Choi–Williams DWT to the noisy data, we obtain the wavelet coefficients,
distribution (CWD) [86] are more precisely in representation of
d ¼ Wyðt i Þ ¼ Wf ðt i Þ þ W ðt i Þ (19)
the energy distribution. Gillespie and Atlas [87] stated that WVD
provides good stationary and non-stationary representation, and Because smaller coefficients are usually contributed by data
provide alternatives for time–frequency analysis in TCM. These noise, thresholding out these coefficients has the effect of
methods are widely applied to mechanical signal analysis for ‘‘removing the data noise.’’ In wavelet thresholding, after setting
machinery condition monitoring [88,89], and shows good time– some coefficients to zeros, the reconstructed (denoised) signal is
frequency energy distribution for fault detection. But in practice obtained by inverse transformation:
these methods are limited by the existence of interference terms,
f^ ðtÞ ¼ W 1 d^ (20)
even if they are attenuated by some approaches [59]. Another
transform similar to WT is the Karhunen–Loeve transform (KLT) Fig. 6 illustrates the thresholding and reconstruction process.
[90]. Tumer et al. [91,92] proposed the KLT to decompose Donoho [71] Donoho and Johnstone [72] developed several
vibration signals from milling into fundamental eigenvectors, wavelet-based thresholding techniques such as hard thresholding
and use the eigenvectors to indicate the changes in the fault and soft thresholding to find an optimal estimate f^ ðtÞ from the
patterns. This is an improvement over methods which project the noisy data. The choice of threshold l is crucial: small/large
data onto pre-determined wave functions, such as FT with sines threshold values will produce estimates that tend to overfit/
and cosines. But this approach is in essence the principal underfit the ffi data. They proposed a universal threshold
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
component analysis (PCA), and it only uses the second-order l ¼ s^ 2 logðnÞ, where n is the number of observations and s^ is
statistics and de-correlates the signals with eigenvector decom- an estimate of the noise variance, which is unknown and needs to
position. It cannot capture non-Gaussian properties (high-order be estimated from noisy samples. Despite the triviality of such a
statistics), which is very important in characterizing signals used threshold, they showed that the resulting wavelet estimator is
in machining monitoring. asymptotically near-minimax among all estimators within the
whole range of the Besov space [100]. Fig. 7 illustrated the
thresholding approaches. The hard thresholding employs a ‘‘keep-
3.2. Wavelet denoising in TCM or-kill’’ rule (Fig. 7(a)), while the soft thresholding is a ‘‘shrink-or-
kill’’ rule (Fig. 7(b)).
Noise always exists in machining, especially encountered in Mathematically speaking, both thresholding approaches have
high precision machining [93–95]. Denoising is a practical drawbacks: hard thresholding is unstable, and sensitive to small
problem in TCM. Houshmand and Elijah Kannatey-Asibu [96] changes, while soft thresholding is a bias estimate rule. To
found that the spectra of the AE signal are highly contaminated by overcome these drawbacks, Gao and Bruce [101] developed the
noise. This noise is so high that the first principal component of firm threshold thresholding. The resulting wavelet thresholding
ARTICLE IN PRESS
wavelet residue probability density

0.5 6
residue
amplitude (N)
Gaussian fit
4
pdf
0
2
-0.5 0
0 0.2 0.4 0.6 0.8 1 -0.3 -0.2 -0.1 0 0.1 0.2 0.3
time (s) sample value
PSD Autocorrelations
0.08
autocorrelation value
0.8
PSD value
0.06 0.6
0.04 0.4
0.2
0.02 0
-0.2
0 500 1000 1500 2000 2500 3000 -200 -100 0 100 200
frequency (Hz) step
Fig. 8. Illustrative of the residue of wavelet denoised force. (a) Estimated noise, (b) noise distribution compared with Gaussian distribution, (c) the corresponding power
spectrum density and (d) the autocorrelation coefficients.
offers a balance between the two approaches. Other ways to reliable to detect a weak signature of mechanical impulse-like
overcome this problem is to estimate level-dependent thresholds defect signal, whereas the wavelet decomposition denoising
with generalized cross-validation [102], or utilize a prior informa- method can achieve satisfactory results on smooth signal
tion with Bayes shrinkage [103,104] for signal denoising, which detection. Lin and Qu [113], and Lin et al. [114] improved the
are exclusively studied in statistics literature and not applied to soft-thresholding method by utilizing a prior information on the
mechanical domain however. (See Jansen [105] for a comprehen- probability density of the impulse, which matched with Morlet
sive study on wavelet denoising.) wavelet. The time–frequency resolution can be adapted to
Menon et al. [106] used the wavelet-based method to eliminate different signals of interest. It was claimed that the method
the background noise, which was a problem when using the AE to performed excellently when applied to denoise gear and bearing
detect small fatigue cracks in rotor head components. Bukkapat- vibration signals with a low SNR.
nam et al. [107] modified Donoho’s thresholding method for However, the existing wavelet denoising methods reported in
chaotic signal with multiplicative noise. It was found that the the literature rely heavily on white Gaussian noise and relative
method can separate chaotic signal from worn tool in machining, energy levels of wavelet coefficients [115]. In practical machining,
and suitable for on-line implementation. The DWT was formerly the noise is generally not Gaussian. This is because the noise is not
applied by the authors as signal separation approach for the purely random but correlated with working conditions. Fig. 8
denoising [108], which was not thresholding, but as a matched shows the residue of the denoised force after soft thresholding.
filter with MRA decomposition. Kwak and Ha [109] described the The PSD is nearly evenly distributed in all frequencies, and a
use of the grinding force signal with noise reduction to detect the Gaussian signal fits it quite well. The wavelet denoising methods
dressing time based on DWT. As a result of denoising, the grinding meet problems in denoising non-Gaussian noise, which is
force signal was successfully used to detect the need for dressing. discussed in details in [98]. It was found that the thresholds
The wavelet denoising method was found to be more effective defined by Donoho [71] Donoho and Johnstone [72] are too small
than the FFT filtering technique. Kwak [110] furthered this when we met super-Gaussian (normalized kurtosis40) noise
approach in turning. DWT is used in both denoising and detecting [116,117]. In this condition, the noise can be regarded as a certain
tool failure. The DWT coefficients of the cutting force signal signal and the wavelet decomposition coefficients are not evenly
showed that the onset time of tool failure and chatter vibration distributed among all scales as Gaussian noise, and as a result
was successfully detected. Li and Guan [111] proposed a wavelet- some of the noise coefficients are not small. Under this condition,
based denoising to extract marked features from the feed-motor independent component analysis (ICA) [118,119] for noise separa-
current signals to indicate the minor cutting edge fracture. It was tion is more effective than wavelet denoising as shown in [98].
found that the best denoising approach was to utilize a third Fig. 9 is the recorded real noise during the interval of machining.
Symmlet mother wavelet function in combination with cross- Apparently, it is quite different from the estimated noise in Fig. 8
validation threshold determination and soft thresholding. both in time and frequency domain, and is super-Gaussian.
Wavelet denoising approaches are also studied in mechanical
system condition monitoring, which is quite similar to the TCM
approach. Qiu et al. [112] compared the performance of wavelet 3.3. Feature extraction and dimension reduction
decomposition-based denoising and wavelet filter-based denois-
ing methods on signals from mechanical defects. The comparison Apart from the original intention of the WT for time–frequency
result demonstrates that wavelet filter is more suitable and analysis, the important and successful application of wavelet in
ARTICLE IN PRESS
noise probability density

0 10
amplitude (N) real noise
Gaussian fit
-0.2
pdf
5
-0.4
0
0 0.2 0.4 0.6 0.8 1 -0.5 -0.4 -0.3 -0.2 -0.1
time (s) sample value
PSD autocorrelation
0.8 1
autocorrelation value
0.6
PSD value
0.9
0.4
0.8
0.2
0 0.7
0 1000 2000 3000 -200 -100 0 100 200
frquency (Hz) step
Fig. 9. Recorded real noise. (a) Noise signal, (b) noise distribution compared to Gaussian distribution, (c) the corresponding power spectrum density and (d) the
autocorrelation coefficients.
Table 4 band. These features are input to MLP for the tool state
Feature extraction with wavelet analysis. classification. It was shown that the features extracted by wavelet
transform are less sensitive to changing cutting conditions and
Features Reference Comments
the MLP has high classification rate. Similar study has also been
DWT [61,67,68,120,122–125,129,130,132] Statistics of DWT coefficients, need reported in [121]. Kamarthi and Pittner [122], and Pittner and
succeeding classifier or set a Kamarthi [123] investigate a flank wear estimation technique in
threshold for classification turning through wavelet representation of acoustic emission
WPD [24,60,83,121,126–128,131,133–136] Statistics of WPD coefficients, need
succeeding classifier or set a
signals. The energies of DWT coefficients are used for flank wear
threshold for classification estimation. The flank wear estimation from the recurrent neural
MP [85] Acted as both feature extraction and network is fairly accurate and indicates that the FWT representa-
classification tion of AE signals is more effective and sensitive than Fourier
transform representation. Kamarti et al. [124] extended this study
with a new feature extraction method based on the FWT. The
features of AE signal are calculated by the Euclidean norms of the
TCM is feature extraction and dimension reduction. The wavelet frequency clusters. They showed that the proposed method can
coefficients usually need to be extracted for robust and effective efficiently extract important features related to progressive flank
representation of different tool states in TCM. The extraction wear. Choi et al. [125] studied cutting force trends and tool wear
approaches are generally based on some statistical measurements effects in ramp cut machining. This study is challenging because
(i.e. mean, variance, Euclidian distance, etc.) to maximize the the depth of cut is continuously changing in ramp cuts. Wavelet
discriminability or classification among different tool states. analysis is applied to cutting forces from a progressively worn
Related literatures are summarized in Table 4 according to their tool. The root mean square value of the approximation coefficients
approaches. extracted as features and linear regression are used for tool wear
Moria et al. [61] developed a method for extracting pre-failure estimation. It is reported that for small depth of cut, the linear
features from the cutting force to predict breakage of a small drill regression can estimate the tool wear with error below 6%.
bit to discriminate different tool states. The DWT coefficients are A typical example of wavelet feature extraction was presented
reduced to three indices: energy, waviness, and irregularity. The by Wu and Du [60]. They introduced an automatic feature
problem is it is difficult to determine the index function for extraction and assessment procedure using a wavelet packet
classification. Tansel et al. [67] used the DWT coefficients of the transform in TCM. To identify the effectiveness of the selected
thrust force as an input to (ART-2) to predict micro drill bit features in both time and frequency domains, four criteria such as
breakage in peck drilling. The authors later enhanced this cross-correlation and cross-coherence of signal and reconstructed
technique by encoding the raw data before inputting it to the signal, correlation of the residue, and power spectrum of the
ART-2 [68]. However, the encoding technique relies excessively on residue, are proposed. The proposed method is tested for chatter
statistical abstractions that cannot be explained in terms of the monitoring in turning and tool wear monitoring in drilling.
machining system behavior during failure. Thus, the drilling It was found that WPD can capture important features of signal
monitoring method lacks a basis for generalization to other types that are sensitive to the machining conditions, e.g. chatter and
of drilling processes. Hong et al. [120] decomposed dynamic different states of tool wear, but is insensitive to the variation of
cutting force signal into different frequency bands by DWT, and working conditions and noises. Accordingly, accurate and reliable
two features, mean values and variances of the local maxima, on-line monitoring decisions can be made. Wu et al. [126]
were extracted from the decomposed signal for each frequency furthered the study with a real-time implementation of WPD
ARTICLE IN PRESS
to motor current signal. The wavelet packet features were 25

obtained under normal tool condition. Based on the principal slight wear
components analysis the alarm thresholds are then constructed 20 medium wear
discriminate value
for decision making. WPD approach for the feature extraction is severe wear
also studied in [127], while the classification is based on a MLP 15

neural network.
Scheffer and Heyns [128] adapted the approach of [60] and 10
used the correlation coefficient with an ideal trend function to
select features during machining. The monitoring system auto- 5
matically selected features displaying the most consistent trend
towards tool wear. The time and frequency domains, time-series 0
coefficients and wavelet packet features were extracted. The WPD 0 5 10 15 20 25 30 35
features consist 5 of 13 features finally selected. Wang et al. [129] the number of wavelet packet features
extracted DWT coefficients and designed a codebook vector to
convert the feature vectors into a symbolic sequence. The Fig. 10. Discriminant properties of different wavelet packets.
effectiveness of the approach was evaluated under different

conditions. The problem of this approach is that the vector 3.4. Singularity analysis for tool wear detection
quantization is targeted for HMM modeling, and may largely
degrade the sensitivity of the wavelet features, since it is averaged In signal analysis, singularity is thought of as either an abrupt
over a long period. Abu-Mahfouz [130] found that the average change or a sudden shift of the signal’s value to a different level.
harmonic wavelet coefficients and the maximum entropy spec- This corresponds to the phenomenon happened in machining
trum peaks are more efficient in training the MLP than the time such as chipping and tool breakage. The original goal of
domain features in monitoring of twist drill wear. Other features, singularity analysis is to estimate the localization and degrees of
such as normalized mean wavelet packet coefficients of the feed a signal’s abrupt changes or image edges [137,138]. It is estimated
force, which is affected not only by the flank wear but also by the from wavelet coefficient regularity at certain point and the rate of
severe crater wear observed in high-speed machining, are tried by decay of the coefficients near this point. It has been studied in
Obikawa and Shinozuka [131]. Li et al. [132] applied DWT to machinery condition monitoring lately [139–147], but little is
decompose turning AE signals into uncorrelated components. studied in TCM. Because it is robust and stable in capturing signal
These components were treated as features and thresholded with changes, it can be expected that these singularity-based features
Shewhart control charts for TCM. The case study demonstrated will provide valuable study in TCM. We introduce the idea of
that the new method is able to detect serious tool wear or wavelet singular analysis first and then reviewed some papers
breakage in the machining process. However, the threshold related to machinery condition monitoring, and then illustrate the
selection is not mentioned and it is difficult to decide the range application to milling TCM.
limit. The singularity of the signal is measured by the Lipschitz
The following case study shows a feature extraction approach exponent (LE) [137]. Assume the signal f(t) has the following
for milling TCM [133]. Milling force is highly non-stationary and property at the vicinity of t0:
WT is more appropriate than FT approaches. The features are
extracted with wavelet packet coefficients. It is decomposed into 5 f ðt 0 þ hÞ Pn ðhÞpAha ; noaon þ 1 (22)
levels for the signal representation. There are overall 62 wavelet
where Pn(t) is an n-order polynomial and pass the point f(t0), and h
packets in this decomposition. These high-dimension features
is small enough, the LE of f(t) at t0 is the superemum of the a for
need to be reduced for less computation and robust representa-
which function (23) holds. It follows directly that a determines
tion. A feature selection algorithm is adapted as Fisher’s
f(t)’s continuously differentiable times (regular) at t0: the higher
discriminant ratio (FDR) [134] to discriminate features from
the a, the more regular the function f(t). Mallat and Hwang [137]
different classes:
have shown that the local regularity of a function (signal) is

c
2 related to the propagation across scale of its wavelet maxima and
X
c X
mi mj
FDR ¼ (21) the decay of the wavelet transform modulus in the time-scale
i jai
s2i þ s2j
plane. A modulus maxima line connects all these local maxima
points.
where the subscripts i, j refer to the mean (mi,mj) and scatter
With the WT of f(t), and in the vicinity of t,
matrix (si,sj) corresponding to the features under investigation for
the classes oi, oj, respectively. Features are then ranked with their
Wf ðtÞpAsa or logWf ðtÞp log A þ a log s (23)
s s
class discriminating ability according to their FRD. Their FDR
score is demonstrated in Fig. 10 under different tool wear level. where A is a constant related to the wavelet c(t). This function
The features are normalized between [0,1] and classified by a connects the wavelet scale j and the Holder exponent a. The A and
succeeding classifier. a can be computed by setting the equality in function (24). The
Besides TCM, interesting and efficient wavelet feature extrac- function also shows the relationship between log wavelet trans-
tion approaches are also studied widely in machinery condition form modulus maxima (WTMM) and the wavelet scale j (or a in
monitoring, which are valuable references to TCM. Wang and Gao CWT). The Holder exponent a indicates the degree of singularity
[135] selected wavelet packet feature with principal component and is often used as a feature in machine condition monitoring.
analysis in machine defect detection. Liu and Ling [136] The singularity analysis with wavelet for TCM was attempted
introduced mutual information to search redundant wavelet by Chen and Li in [139]. The authors discussed the WTMM
packets. This algorithm extracts those informative wavelets which approach but it was not implemented as they claimed that the
carry important but less redundant information about machinery WTMM was easily influenced by the environment noise and
faults. It was found that the proposed feature extraction method varying cutting conditions; it could not represent the cutting tool
performed much better than the PCA in detecting diesel engine conditions accurately. Therefore, the WTMM approach was not
malfunctions. preferred but chose the wavelet coefficient norm and their
ARTICLE IN PRESS
statistics as features to indicate the tool condition instead. This can be used as a stable indication of different gear faults. It was
approach is more robust in case of noise but it is also subject to also showed that the Holder exponent was effective in capturing
the limitation in choosing threshold, which also varies with non-stationary nature of signals and sensitive for identifying
working conditions. Besides, the AE signal can be denoised first to structure damage [145]. Peng et al. [146] examined shaft orbits
apply the WTMM scheme. using the wavelet modulus maxima. Four LE-based features are
Most applications of wavelet for singularity detection were extracted to classify the shaft orbit. Peng et al. [147] furthered
applied such as change point analysis [140], and non-linear this study and presented a novel singularity-based fault feature
system identification [141], which are similar to the TCM in extraction from vibration signals. The data are collected
machining. Hambaba and Hu [142] used wavelet transform to under different machine health conditions that show different
determine the Holder exponent value of a gear response at patterns of singularities measured quantitatively by the
different scale levels. By fitting an autoregressive moving average Lipschitz exponents. The number of Lipschitz exponents per
(ARMA) model to the wavelet-transformed data, analysis of the rotation, the mean value and the relative standard deviation
residual error was used to indicate the presence of damage in the of the Lipschitz exponents that are obtained from the extracted
gear. Sun et al. [143] proposed singularity analysis for bearing features for singularity representation. The results show that the
defect diagnosis. Through modifying the intensity of the wavelet three parameters are excellent fault features for fault pattern
transform modulus maxima, defect-related vibration signature recognition.
was highlighted and could be easily associated with the bearing Fig. 11 shows an example study with modulus maxima for
defect characteristic frequencies. Loutridis and Trochidis [144] detection of milling tool at different wear state. Fig. 11a shows a
employed the Lipschitz exponent to investigate gear faults. It was cutting force signal from a fresh tool and severe worn tool.
shown that the Hoelder exponent for each type of fault exhibits a Lipschitz exponent extracted from the WTMM (Fig. 11(b)). Though
constant value, not affected by changing working conditions, and they do not change much, Lipschitz exponents offer a stable tool
Slight wear Severe wear

2 4
force amplitude (N)
force amplitude (N)
1
2
0
0
-1
-2 -2
0 500 1000 1500 2000 2500 3000 3500 4000 4500 0 500 1000 1500 2000 2500 3000 3500 4000 4500
force sample force sample
Holder estimates at time instants T Holder estimates at time instants T

0.5 1
0 0.5
HE value
HE value
-0.5 0
-1 -0.5
-1.5 -1
0 500 1000 1500 2000 2500 3000 3500 4000 4500 0 500 1000 1500 2000 2500 3000 3500 4000 4500
sample force sample
Singularity Spectrum Singularity Spectrum

1 1
0.95
0.9
0.95
spectrum value
spectrum value
0.85
0.8
0.9 0.75
0.7
0.65
0.85
0.6
0.55
0.8 0.5
-0.85 -0.8 -0.75 -0.7 -0.65 -0.3 -0.25 -0.2 -0.15 -0.1 -0.05 0 0.05 0.1
Singularity α Singularity α
Fig. 11. (a) Cutting forces of fresh tool and severe flank wear tool. (b) Lipschitz exponent of slight wear tool and severe wear tool. (c) The singular spectrum of the slight
wear tool and severe wear tool. Modulus maxima for slight wear tool and severe wear tool.
ARTICLE IN PRESS
to detect the degrees of tool wear level as shown. The spectrum of [149] present an application of wavelets analysis to density
LEs from two tool states is illustrated in Fig. 11(c). It is observed estimation and process monitoring. The resulting density function
that with the fault severity increasing, the milling force signals’ was used to define a normal operating region for the process so
singularities and singularity ranges will increase as well, and that any future abnormal changes in the process can be
therefore we can evaluate the fault severity through measuring monitored. Results of applying these techniques to chemical
the force signals’ singularities and singularity ranges. process are also presented to be effective in monitoring the states.
Carefully attention should also be made upon choosing wavelet In the structural health monitoring, Krishnan and Kiremidjian
in the WTMM approach. Not all wavelet’s WTMM location [150] applied a time-series-based detection algorithm with
corresponds to the abrupt change points, and it happens only GMMs. The structural vibration signals are modeled with ARMA
when the applied wavelet is the first-order differentiation of a first, and then the first three AR coefficients are chosen as feature
certain smooth function (i.e. Gaussian). If it is the second-order vectors. A Gaussian mixture model is used to model the feature
differentiation of a smooth function, the change point will be the vector. Damage is detected by comparing the Mahalanobis
WT’s zero crossing position, not the WTMM position. As in this distance of the extracted AR coefficients and those of the
study it applies the first-order differentiation of Gaussian function undamaged. This approach provides a useful framework for data
in this study of Fig. 11. The regularity is another important fusion, where different measurements such as strains, tempera-
criterion in the selection wavelet. Usually, the selected wavelet ture, and humidity could be used for a more robust damage
must be sufficiently regular, which means a larger vanish moment decision.
(see [65] for details about vanish moment); otherwise some The most important and popular density estimation approach
singularities would be overlooked. Additionally, the noise will before wavelet density estimation is kernel density estimation
influence the performance of the wavelet greatly, therefore before [151]. Wavelet is brought attention in density estimation lately
the singularity is detected, the signal preprocessing must be [152]. The differences between kernel and wavelet estimates are
carried out. mainly explained by the ability of the wavelet method to take into
account local gaps in the data distribution. This new approach is
very promising, since smaller structures superimposed onto a
3.5. Wavelet probability density estimation for tool larger one are detected by this technique, especially when small
state classification samples are investigated. Thus, wavelet solutions appear to be
better suited for classification studies.
Probability density estimation is a classic problem in statistics. Wavelet density estimation is developed from non-parametric
The density estimation problem in statistics is in fact the pattern regression problem in statistics example [153,154]. Wavelet
recognition problem in engineering as discussed in Section 1.2: density estimation opens the problem of density estimation for
assigning classes to higher probability density groups. There is dependent observations, while almost all other density estimators
little study of probability density application to TCM except fail in dealing. Let X1,X1,y,Xn be a sequence of independently and
[47,148]. Heck and Chou [47], and Chou and Heck [148] reported identically distributed random variables, with density: h ¼ h(x).
R
studies of wavelet density estimation for TCM. They use a This density is unknown and to be estimate. Suppose h(x)2 dx is
Gaussian mixture model to approximate the wavelet coefficients finite, and approximated in wavelet series:
from machining signals. Different tool states are represented with
different GMM, and then choose the maximum class conditional X J
X X
probability for state decision when feature extracted for classifi- hðxÞ ¼ cj;k fj;k ðtÞ þ dJ;k cJ;k ðtÞ (24)
cation. In other area as chemical process monitoring, Safavi et al. k2Z j¼1 k2Z
10 6
8
4
6
p(x)
p(x)
4 2
2
0
0
-2 -2
-1 -0.5 0 0.5 1 1.5 -1 -0.5 0 0.5 1 1.5 2
x x
2 0.5
0.4 Pmax
1 0.3
p(x)
0.2
0
0.1
0 Pmin
-1
-2 -1 0 1 2 3 4 0 5 10 15 20 25
x
Fig. 12. Density estimation with wavelet and state estimation. (a) slight wear pdf, (b) medium wear pdf, (c) severe wear pdf and (d) Bayesian classifier (MAP).
ARTICLE IN PRESS
To estimate h, it is sufficient to estimate the coordinates cj,k and be very helpful in TCM. For future studies, the wavelet density
dJ,k by the sample mean: estimation can approximate wider densities than other density
estimation approaches. The density estimation actually incorpo-
1X n
1 X
n
c^ j;k ¼ f ðX Þ; d^ J;k ¼ cJ;k ðX i Þ (25) rates both feature extraction and classification steps in the TCM
n i¼1 j;k i n i¼1 and is concise and easy to implement. More studies can be
Inserting this estimate to the wavelet series expansion of h, we expected in these areas.
get an estimate
J
References
X X X
^
hðxÞ ¼ c^ j;k fj;k ðtÞ þ d^ J;k cJ;k ðtÞ (26)
k2Z j¼1 k2Z [1] Milton C. Shaw, Metal Cutting Principles, second ed., Oxford University Press,
New York, 2005.
^
The estimate hðxÞ is then used as real probability density for [2] Geoffrey Boothroyd, A. Winston, Knight, Fundamentals of Machining and
application. Machine Tools, Addison-Wesley, Redwood City, CA, 1989.
[3] G. Byrne, D. Dornfeld, I. Inasaki, G. Ketteler, W. Konig, R. Teti, Tool condition
The idea can be easily adapt to tool wear classification with the monitoring (TCM)—the status of research and industrial application, Annals
following steps: of the CIRP 44 (1995) 541.
[4] D.E. Dimla, Sensor signal for tool-wear monitoring in metal cutting
operations—a review of methods, International Journal of Machine Tools
(1) Estimate the respective density distributions of tool wear and Manufacture 40 (2000) 1073–1098.
levels provided by the machining data and measured tool [5] S.Y. Liang, L.H. Rogelio, G.L. Robert, Machining process monitoring and
wear, like the supervised neural network training. control: the state-of-the-art, Journal of Manufacturing Science and En-
gineering 126 (2004) 297–310.
(2) Match the present estimated signal density to the known [6] D.C.D. Oguamanam, H. Raafat, S.M. Taboun, A machine vision system for
density. wear monitoring and breakage detection of single-point cutting tools,
(3) Select the maximum posterior probability density as tool Computers in Industrial Engineering 26 (3) (1994) 575–598.
[7] S. Kurada, C. Bradley, A review of machine vision sensors for tool condition
condition with Bayes classifier.
monitoring, Computers in Industry 34 (1) (1997) 55–72.
[8] W.H. Wang, G.S. Hong, Y.S. Wong, K.P. Zhu, Sensor fusion for on-line tool
Fig. 12(a)–(c) is the learned density from the measurement, condition monitoring in milling, International Journal of Production
Research 45 (21) (2007) 5059–5116.
representing three possible tool state. The estimated density is [9] Y. Altintas, In-process detection of tool breakages using time series
matched to these tree conditions, and the likelihood of these monitoring of cutting forces, International Journal of Machine Tools and
matching is found in Fig. 12(d). We conclude that the state highest Manufacture 28 (2) (1988) 157–172.
[10] Y. Altintas, I. Yellowley, In process detection of tool failure in milling using
probability is what we estimate. cutting force models, Journal of Engineering for Industry—Transactions of
The wavelet density estimation generally applied together with the ASME 111 (1989) 149–157.
GMM because this approach is more flexible since it can [11] M.A. Elbestawi, T.A. Papazafiriou, R.X. Du, In-process monitoring of tool wear
in milling using cutting force signature, International Journal of Machine
approximate wider densities. This idea is similar to those of NN, Tools and Manufacture 31 (1) (1991) 55–73.
HMM, and PR approaches in nature, where the approaches find [12] R. Du, M.A. Elbestawi, S.M. Wu, Computer automated monitoring of
the most probable state by the pdfs they belong to. Wavelet manufacturing processes: Part 1. Monitoring decision-making methods,
Journal of Engineering for Industry—Transactions of the ASME May (1995).
density estimation is more powerful as it can approximate wider
[13] R. Du, M.A. Elbestawi, S.M. Wu, Computer automated monitoring of
densities while PR approaches such as LDA assume Gaussian manufacturing processes: Part 2. Applications, Journal of Engineering for
density as a prior, the later is generally not true of machining Industry—Transactions of the ASME May (1995).
[14] H. Saglam, A. Unuvar, Tool condition monitoring in milling on cutting forces
sensory signals however. More studies can be expected in TCM
by a neural network, International Journal of Production Research 41 (7)
with this approach. (2003) 1519–1532.
[15] G. Sutter, A. Molinari, Analysis of the cutting force components and friction
in high speed machining, Journal of Manufacturing Science and Engineer-
4. Conclusion and future studies ing—Transactions of the ASME 127 (May) (2005) 245–250.
[16] J.C. Chen, W.L. Chen, A tool breakage detection system using an accel-
erometer sensor, Journal of Intelligent Manufacturing 10 (2) (1999) 187–197.
It is shown that WT is competitive to other signal analysis [17] T.I. El-Wardany, D. Gaoand, M.A. Elbestawi, Tool condition monitoring in
drilling using vibration signature analysis, International Journal of Machine
approaches because of its MRA, sparsity, and localization proper-
Tools and Manufacture 36 (6) (1996) 687–711.
ties. It is effective in analyzing non-stationary machining sensor [18] D.E. Dimla Sr., P.M. Lister, On-line metal cutting tool condition monitoring I:
signals. Based on the benefits of WT discussed, applications of force and vibration analyses, International Journal of Machine Tools and
wavelet in TCM are reviewed in time–frequency analysis, denois- Manufacture 40 (2000) 739–768.
[19] D.E. Dimla Sr., P.M. Lister, On-line metal cutting tool condition monitoring II:
ing, feature extraction, singularity analysis, and density estima- tool-state classification using multi-layer perceptron neural networks,
tion. It achieves a lot of success in TCM overall but many more International Journal of Machine Tools and Manufacture 40 (2000) 769–781.
need to be further studied. [20] B.S. Berger, I. Minis, J. Harley, M. Rokni, M. Papadopoulos, Wavelet based
cutting state, identification, Journal of Sound and Vibration 213 (5) (1998)
As discussed, the WT presents a natural MRA analysis and 813–827.
keeps the quality factor constant under different scales. These are [21] E.N. Diei, D.A. Dornfeld, A model of tool fracture generated acoustic emission
to the benefit of characterizing signal’s crossing scales properties during machining, Journal of Engineering for Industry—Transactions of the
ASME 109 (1987) 227–233.
with WT. These advantages of WT should be taken over other [22] M. Liu, S.Y. Liang, Analytical modeling of acoustic emission for monitoring of
methods such as FT and PCA in the applications but not only use peripheral milling process, International Journal of Machine Tools and
WT as filter banks and decompose the signal into various Manufacture 31 (4) (1991) 589–606.
[23] AliA Houshmand, Elijah Kannatey-Asibu Jr., Gray D. Herrin, A dynamic model
frequency bands. For the wavelet denoising, the general thresh- for tool detection using acoustic emission, Mechanical Systems and Signal
olding approaches are not preferred in TCM because of its Processing 9 (4) (1995) 415–428.
Gaussian noise assumption, because in TCM the noise is generally [24] Y.M. Niu, Y.S. Wong, G.S. Hong, Multi-category classification of tool
conditions using wavelet packets and ART2 network, Journal of Manufactur-
correlated to machining conditions and not Gaussian. As reference
ing Science and Technology—Transactions of the ASME 120 (1998) 807–815.
noise signal can always be collected during TCM, apply Bayesian [25] P. Srinivasa Pai, P.K. Ramakrishna Rao, Acoustic emission analysis for tool
approaches that incorporate this a prior information will be quite wear monitoring in face milling, International Journal of Production
helpful in signal denoising. WT has great potential in detecting Research 40 (5) (2002) 1081–1093.
[26] J. Sun, G.S. Hong, M. Rahman, Y.S. Wong, Identification of feature set for
abrupt changes of tool conditions in TCM. It is very robust as the effective tool condition monitoring by acoustic emission sensing, Interna-
LE component is insensitive to changing working conditions and is tional Journal of Production Research 42 (5) (2004) 901–918.
ARTICLE IN PRESS
[27] Y. Altintas, Prediction of cutting forces and tool breakage in milling from [57] D. Gabor, Theory of communication, Journal of the Institutional Electrical
feed drive current measurement, Journal of Engineering for Industry— Engineering 93 (III) (1946) 429–457.
Transactions of the ASME 114 (1992) 386–392. [58] M.J. Dowling, Application of non-stationary analysis to machinery monitor-
[28] X. Li, On-line detection of the breakage of small diameter drills using current ing, ICASSP-93, IEEE International Conference on Acoustics, Speech, and
signature wavelet transform, International Journal of Machine Tools and Signal Processing 1 (1993) 59–62.
Manufacture 39 (1999) 157–164. [59] S.G. Mallat, A Wavelet Tour of Signal Processing, second ed., Academic Press,
[29] X. Li, S.K. Tso, Jun Wang, Real time tool condition monitoring using wavelet New York, 1999.
transforms and fuzzy techniques, IEEE Transactions on Systems, Man, and [60] Ya Wu, R. Du, Feature extraction and assessment using wavelet packets for
Cybernetics Part C: Applications and Reviews 31 (3) (2000) 352–357. monitoring of machining process, Mechanical Systems and Signal Processing
[30] X. Li, Detection of tool flute fracture in end milling using feed-motor current 10 (1) (1996) 29–53.
signatures, IEEE/ASME Transactions on Mechatronics 6 (2001) 491–498. [61] K. Mori, N. Kasashima, J.C. Fu, K. Muto, Prediction of small drill bit breakage
[31] B.Y. Lee, Y.S. Tarng, Application of the discrete wavelet transform to the by wavelet transforms and linear discriminant functions, International
monitoring of tool failure in end milling using the spindle motor current, Journal of Machine Tools and Manufacture 39 (1999) 1471–1484.
International Journal of Advanced Manufacturing Technology 15 (1999) [62] W. Gong, T. Obikawa, T. Shirakashi, Monitoring of tool wear states in turning
238–243. based on wavelet analysis, JSME International Journal (Series C) 40 (1997)
[32] H.K. Tonshoff, X. Li, C. Lapp, Application of fast Haar transform and 447–453.
concurrent learning to tool-breakage detection in milling, IEEE/ASME [63] M.C. Yoon, D.H. Chin, Cutting force monitoring in the end milling operation
Transactions on Mechatronics 8 (2003) 414–417. for chatter detection, IMechE 2005 Proceedings IMechE, vol. 219, Part B:
[33] C.S. Leem, D.A. Dornfeld, Design and implementation of sensor based tool Journal of Engineering Manufacture, pp. 455–465.
wear monitoring systems, Mechanical Systems and Signal Processing 10 (4) [64] S.G. Mallat, A theory of multiresolution signal decomposition: the wavelet
(1996) 439–458. representation, IEEE Transactions on Pattern and Machine Intelligence 11 (7)
[34] A. Ghasempoor, J. Jeswiet, T.N. Moore, Real time implementation of on-line (1989) 674–693.
tool condition monitoring in turning, International Journal of Machine Tools [65] I. Daubechies, The wavelet transformation, time–frequency localization
and Manufacture 39 (12) (1999) 1883–1902. and signal analysis, IEEE Transactions on Information Theory 36 (1990)
[35] T.H.C. Childs, K. Maekawa, T. Obikawa, Y. Yamane, Metal Machining: Theory 961–1005.
and Applications, Elsevier, Amsterdam, 2000. [66] I. Daubeches, Ten Lectures on Wavelets, SIAM, Philadelphia, PA, 1992.
[36] Richard O. Duda, Peter E. Hart, David G. Stork, Pattern Classification, second [67] I.N. Tansel, C. Mekdeci, O. Rodriguez, B. Uragun, Monitoring drill conditions
ed., Wiley Interscience, New York, 2001. with wavelet based encoding and neural networks, International Journal of
[37] S.Y. Liang, D.A. Dornfeld, Detection of cutting tool-wear using adaptive time Machine Tools and Manufacture 33 (1993) 559–575.
series modeling of acoustic emission signal, Sensors for Manufacturing— [68] I.N. Tansel, C. Mekdeci, C. McLaughlin, Detection of tool failure in end
Transactions of the ASME (1987) 27–38. milling with wavelet transformations and neural networks (WT-NN),
[38] S.A. Kumar, H.V. Ravindra, Y.G. Srinivasa, In-process tool wear monitoring International Journal of Machine Tools and Manufacture 35 (August)
through time series modeling and pattern recognition, International Journal (1995) 1137–1147.
of Production Research 35 (3) (1997) 739–751. [69] M.V. Wickerhauser, R.R. Coifman, Entropy based methods for best
[39] Bernhard Sick, Review: on-line and indirect tool wear monitoring in turning basis selection, IEEE Transactions on Information Theory 38 (2) (1992)
with artificial neural networks: a review of more than a decade of research, 719–746.
Mechanical Systems and Signal Processing 16 (4) (2002) 487–546. [70] Z.K. Peng, F.L. Chu, Review: application of the wavelet transform in machine
[40] R.G. Silva, R.L. Reuben, K.J. Baker, S.J. Wilcox, Tool wear monitoring of turning condition monitoring and fault diagnostics: a review with bibliography,
operations by neural network and expert system classification of a feature Mechanical Systems and Signal Processing 18 (2004) 199–221.
set generated from multiple sensors, Mechanical Systems and Signal [71] D.L. Donoho, De-noising by soft-thresholding, IEEE Transactions on Informa-
Processing 12 (2) (1998) 319–332. tion Theory 41 (3) (1995) 613–627.
[41] J.H. Lee, D.E. Kim, S.J. Lee, Application of neural networks to flank [72] D.L. Donoho, I.M. Johnstone, Adapting to unknown smoothness via wavelet
wear prediction, Mechanical Systems and Signal Processing 10 (3) (1996) shrinkage, Journal of the American Statistical Association 90 (432) (1995)
265–276. 1200–1224.
[42] H. Saglam, A. Unuvar, Tool condition monitoring in milling based on cutting [73] Isa Yesilyurt, End mill breakage detection using mean frequency analysis of
forces by a neural network, International Journal of Production Research 41 scalogram, International Journal of Machine Tools and Manufacture 46 (3–4)
(7) (2003) 1519–1532. (2006) 450–458.
[43] A. Ghasempoor, J. Jeswiet, T.N. Moore, Real time implementation of on-line [74] M.K. Khraisheh, C. Pezeshki, A.E. Bayoumi, Time series based analysis for
tool condition monitoring in turning, International Journal of Machine Tools primary chatter in metal cutting, Journal of Sound and Vibration 180 (1)
and Manufacture 39 (1999) 1883–1902. (1995) 67–87.
[44] S.C. Lin a, R.J. Lin, Tool wear monitoring in face milling using force signals, [75] Peter W. Tse, Wen-Xian Yang, H.Y. Tam, Machine fault diagnosis through an
Wear 198 (1996) 136–142. effective exact wavelet analysis, Journal of Sound and Vibration 277 (2004)
[45] S. Li, M.A. Elbestawi, Fuzzy clustering for automated tool condition 1005–1024.
monitoring in machining, Mechanical Systems and Signal Processing 10 [76] X. Li, S. Dong, Z. Yuan, Discrete wavelet transform for tool breakage
(5) (1996) 533–550. monitoring, International Journal of Machine Tools and Manufacture 39
[46] Xiaoli Li, Yuan Zhejun, Tool wear monitoring with wavelet packet transform- (1999) 1935–1944.
fuzzy clustering method, Wear 219 (2) (1998) 145–154. [77] J.C. Fu, C.A. Troy, K. Mori, The prediction of small drill bit breakage by
[47] Larry P. Heck, Kenneth C. Chou, Gaussian mixture model classifiers for wavelet based joint time-scale analysis, CIRP Proceedings—Manufacturing
machine monitoring, ICASSP-94, IEEE International Conference on Acoustics Systems 25 (1996) 385–389.
Speech and Signal Processing 6 (6) (1994) VI/133–VI/136. [78] C.S. Suh, P.P. Khurjekar, B. Yang, Characterisation and identification of
[48] R. Kothamasu, S.H. Huang, W.H. Verduin, Comparison of computational dynamic instability in milling operation, Mechanical Systems and Signal
intelligence and statistical methods in condition monitoring for hard Processing 16 (5) (2002) 853–872.
turning, International Journal of Production Research 43 (3) (2005) 597–610. [79] B.S. Berger, I. Minis, J. Harley, M. Rokni, M. Papadopoulos, Wavelet based
[49] Lane M.D. Owsley, Les E. Atlas, Gary D. Bernard, Self-organizing feature maps cutting state identification, Journal of Sound and Vibration 213 (5) (1998)
and hidden Markov models for machine-tool monitoring, IEEE Transactions 813–827.
on Signal Processing 45 (11) (1997) 2787–2796. [80] F. Klocke, M. Reuber, H. Kratz, Application of a wavelet-based signal analysis
[50] Huseyin Metin Ertunc, Cuneyt Oysu, Drill wear monitoring using cutting for evaluating the tool state in cutting operations, Industrial Electronics
force signals, Mechatronics 14 (2004) 533–548. Society, IECON 2000, in: Twenty-sixth Annual Conference of the IEEE, vol. 3,
[51] P. Baruah, R.B. Chinnam, HMMs for diagnostics and prognostics in machining 22–28 October 2000, pp. 1967–1972.
processes, International Journal of Production Research 43 (6) (2005) [81] F.G. Luis Alfonso, H.R. Gilberto, P.V. Rocio, R.T. Rene de Jesus, L.T. Wbaldo,
1275–1293. Sensorless tool failure monitoring system for drilling machines, Interna-
[52] J. Sun, G.S. Hong, Y.S. Wong, M. Rahman, Z.G. Wang, Effective training data tional Journal of Machine Tools and Manufacture 46 (2006) 381–386.
selection in tool condition monitoring system, International Journal of [82] Y.M. Niu, Information-driven tool condition monitoring techniques, Ph.D.
Machine Tools and Manufacture 46 (2006) 218–224. Thesis, Department of Mechanical Engineering, Faculty of Engineering,
[53] Z.J. Pasek, Exploration of rough sets theory use for manufacturing process National University of Singapore, 2001.
monitoring, IMechE 2006 Proceedings ImechE, vol. 220, Part B: Journal of [83] B. Liu, Selection of wavelet packet basis for rotating machinery fault
Engineering Manufacture, pp. 335–374. diagnosis, Journal of Sound and Vibration 284 (2005) 567–582.
[54] Alan V. Oppenheim, Alan S. Willsky, S. Nawab Nawab, Syed ham [84] S.G. Mallat, Z. Zhang, Matching pursuits with time–frequency dictionaries,
Nawab, Signals and Systems, second ed., Prentice-Hall, Englewood Cliffs, IEEE Transactions on Signal Processing 41 (1993) 3397–3415.
NJ, 1996. [85] J.C. Fu, C.A. Troy, P.J. Phillips, A matching pursuit approach to small drill bit
[55] Y.S. Tarng, B.Y. Lee, A sensor for the detection of tool breakage in NC milling, breakage prediction, International Journal of Production Research 37 (14)
Journal of Materials Processing Technology 36 (1993) 259–272. (1999) 3247–3261.
[56] Rixin Zhu, Shiv G. Kapoor, Richard E. DeVor, A model-based monitoring and [86] L. Cohen, Time–frequency distribution—a review, Proceedings of the IEEE 77
fault diagnosis methodology for free-form surface machining process, (1989) 941–981.
Journal of Manufacturing Science and Engineering—Transactions of the [87] B.W. Gillespie, L.E. Atlas, Optimizing Time–Frequency Kernels for Classifica-
ASME 125 (August) (2003). tion, IEEE Transactions on Signal Processing 49 (3) (2001) 485–496.
ARTICLE IN PRESS
[88] S.K. Lee, P.R. White, Higher-order time–frequency analysis and its applica- [117] R. Averkamp, C. Houdr, Wavelet thresholding for non-necessarily Gaussian
tion to fault detection in rotating machinery, Mechanical Systems and Signal noise: idealism, Annals of Statistics 31 (2003) 110–151.
Processing 11 (4) (1997) 637–650. [118] J.F. Cardoso, Blind signal separation: statistical principles, Proceedings of the
[89] W.J. Staszewski, K. Worden, G.R. Tomlinson, Time–frequency analysis in IEEE 9 (10) (1998) 2009–2025.
gearbox fault detection using the Wigner–Ville distribution and pattern [119] A. Hyvarinen, J. Karhunen, E. Oja, Independent Component Analysis, Wiley,
recognition, Mechanical Systems and Signal Processing 11 (5) (1997) New York, 2001.
673–692. [120] G.S. Hong, M. Rahman, Q. Zhou, Using neural networks for tool condition
[90] Michael Kirby, Geometric Data Analysis: An Empirical Approach to monitoring based on wavelet decomposition, International Journal of
Dimensionality Reduction and the Study of Patterns, Wiley-Interscience, Machine Tools and Manufacturing (1996) 36–55.
New York, 2000. [121] Q. Zhou, G.S. Hong, M. Rahman, A new tool life criterion for tool condition
[91] I. Tumer, K. Wood, I. Busch-Vishniac, 1997, Monitoring fault condition during monitoring using a neural network, Engineering Applications of Artificial
manufacturing using the Karhunen–Loeve transform, in: ASME Mechanical Intelligence 8 (5) (1995) 579–588.
Vibration and Noise Conference, System Health Monitoring Symposium, [122] S.V. Kamarthi, S. Pittner, Fast Fourier and wavelet transform for flank wear
vol. DETC97-VIB4234. estimation—a comparison, Mechanical Systems and Signal Processing 11
[92] I.Y. Tumer, K.L. Wood, I.J. Busch-Vishniac, Monitoring of signals from (November) (1997) 791–809.
manufacturing processes using the Karhunen–Loee Ve transform, Mechan- [123] S. Pittner, V. Kamarthi, Feature extraction from wavelet coefficients for
ical Systems and Signal Processing 14 (6) (2000) 1011–1026. pattern recognition tasks, IEEE Transactions on Pattern Analysis and
[93] I.N. Tansel, O. Rodriguez, M. Trujillo, E. Paz, W. Li, Micro-end-milling—I. Machine Intelligence 21 (1) (1999) 83–88.
Wear and breakage, International Journal of Machine Tools and Manufacture [124] S.V. Kamarti, S.R.T. Kumara, P.H. Cohen, Flank wear estimation in turning
38 (1998) 1419–1436. through wavelet representation of acoustic emission signals, Journal of
[94] I.N. Tansel, T.T. Arkan, W.Y. Bao, N. Mahendrakar, B. Shisler, D. Smith, M. Manufacturing Science and Engineering—Transactions of the ASME 122
McCool, Tool wear estimation in micro-machining. Part II: neural-network- (2000) 12–19.
based periodic inspector for nonmetals, International Journal of Machine [125] Y. Choi, R. Narayanaswami, A. Chandra, Tool wear monitoring in ramp cuts in
Tools and Manufacture 40 (2000) 609–620. end milling using the wavelet transform, International Journal of Advanced
[95] K.P. Zhu, Y.S. Wong, G.S. Hong, Noise-robust tool condition monitoring in Manufacturing Technology 23 (2004) 419–428.
micro-milling with hidden Markov models, in: P. Bhanu (Ed.), Soft [126] Ya Wu, Philippe Escande, R. Du, A new method for real-time tool condition
Computing Applications in Industry, Studies in Fuzziness and Soft Comput- monitoring in transfer machining stations, Journal of Manufacturing Science
ing 226, pp. 23–46, Springer, 2008. and Engineering 123 (May) (2001) 339–347.
[96] Ali A. Houshmand, J.R. Elijah Kannatey-Asibu, Statistical process control for [127] S.T.S. Bukkapatnam, S.R.T. Kumara, A. Lakhtakia, Analysis of acoustic
cutting tool of acoustic emission monitoring, Mechanical Systems and Signal emission in machining, Journal of Manufacturing Science and Engineer-
Processing 3 (4) (1989) 405–424. ing—Transactions of the ASME 121 (1999) 568–576.
[97] P.T. Huang, J.C. Chen, C.Y. Chou, A statistical approach in detecting tool [128] C. Scheffer, P.S. Heyns, Wear monitoring in turning operations using
breakage in end milling operations, Journal of Industrial Technology 15 (3) vibration and strain measurements, Mechanical Systems and Signal
(1999). Processing 15 (6) (2001) 1185–1202.
[98] K.P. Zhu, G.S. Hong, Y.S. Wong, W.H. Wang, Cutting force denoising in micro- [129] Litao Wang, Mostafa G. Mehrabi, Elijah Kannatey-Asibu, Hidden Markov
milling tool condition monitoring, International Journal of Production model-based tool wear monitoring in turning, Journal of Manufacturing
Research 46 (16) (2008) 4391–4408. Science and Engineering—Transactions of the ASME 124 (August) (2002)
[99] M. Jansen, Wavelet denoising, Ph.D. Thesis, Department of Computer 652–658.
Science, Katholieke Universiteit Leuven, Belgium, October 2001. [130] I. Abu-Mahfouz, Drilling wear detection and classification using vibration
[100] D.L. Donoho, I.M. Johnstone, Minimax estimation via wavelet shrinkage, signals and artificial neural network, International Journal of Machine Tools
Annals of Statistics 26 (1998) 879–921. and Manufacture 43 (7) (2003) 707–720.
[101] H.Y. Gao, A.G. Bruce, WaveShrink with firm shrinkage, Statistica Sinica 7 [131] Toshiyuki Obikawa, Jun Shinozuka, Monitoring of flank wear of coated tools
(1997) 855–874. in high speed machining with a neural network ART2, International Journal
[102] G.P. Nason, Wavelet shrinkage using cross-validation, Journal of the Royal of Machine Tools and Manufacture 44 (2004) 1311–1318.
Statistical Society B 58 (1996) 463–479. [132] X. Li, Y. Yao, Z. Yuan, On-line tool condition monitoring system with wavelet
[103] H.A. Chipman, E.D. Kolaczyk, R.E. McCulloch, Adaptive Bayesian wavelet fuzzy neural network, Journal of Intelligent Manufacturing 8 (1997)
shrinkage, Journal of the American Statistical Association 92 (1997) 271–276.
1413–1421. [133] K.P. Zhu, Y.S. Wong, G.S. Hong, Multi-category micro-milling tool wear
[104] F. Abramovich, T. Sapatinas, Bayesian approach to wavelet decomposition classification with continuous hidden Markov models, Mechanical Systems
and shrinkage, in: P. Müller, B. Vidakovic (Eds.), Bayesian Inference in and Signal Processing 23 (2009) 547–560.
Wavelet Based Models, Lecture Notes in Statistics, vol. 141, Springer, [134] S. Theodoridis, K. Koutroumbas, Pattern Recognition, Academic Press, New
New York, 1999, pp. 33–50. York, 2003.
[105] M. Jansen, Noise Reduction by Wavelet Thresholding, Lecture Notes in [135] C. Wang, R.X. Gao, Wavelet transform with spectral post-processing for
Statistics, vol. 161, Springer, Berlin, 2001. enhanced feature extraction, IEEE Transactions on Instrumentation and
[106] S. Menon, J.N. Schoess, R. Hamza, D. Busch, Wavelet-based acoustic emission Measurement 52 (2003) 1296–1301.
detection method with adaptive thresholding, Proceedings of the SPIE 3986 [136] B. Liu, S.F. Ling, On the selection of informative wavelets for machinery
(2000) 71–77. diagnosis, Mechanical Systems and Signal Processing 13 (1) (1999) 145–162.
[107] S.T.S. Bukkapatnam, S.R.T. Kumara, A. Lakhtakia, P. Srinivasan, The [137] S.G. Mallat, W.L. Hwang, Singularity detection and processing with wavelets,
neighborhood method and its coupling with the wavelet method for IEEE Transactions on Information Theory 38 (1992) 617–643.
signal separation of chaotic signals, Signal Processing 82 (2002) 1351–1374. [138] S. Mallat, S. Zhong, Characterization of signals from multiscale edges, IEEE
[108] S.T.S. Bukkapatnam, Soundar R.T. Kumara, A. Lakhtakia, Fractal estimation of Transactions on Pattern Analysis and Machine Intelligence 14 (7) (1992).
flank wear in turning, Journal of Dynamic Systems, Measurement, and [139] Xiaozhi Chen, Beizhi Li, Acoustic emission method for tool condition
Control—Transactions of the ASME 122 (March) (2000) 89–94. monitoring based on wavelet analysis, International Journal of Advanced
[109] Jae-Seob Kwak, Man-Kyung Ha, Detection of dressing time using the Manufacturing Technology 33 (2007) 968–976.
grinding force signal based on the discrete wavelet decomposition, The [140] R.T. Ogden, J.D. Lynch, Bayesian analysis of change-point models, Lecture
International Journal of Advanced Manufacturing Technology 23 (1–2) Notes in Statistics, vol. 141, 1999, pp. 67–82.
(2004) 87–92. [141] W.J. Staszewski, Identification of non-linear systems using multi-scale ridges
[110] Jae-Seob Kwak, Application of wavelet transform technique to detect tool and skeletons of the wavelet transform, Journal of Sound and Vibration 214
failure in turning operations, The International Journal of Advanced (4) (1998) 639–658.
Manufacturing Technology 28 (11–12) (2006) 1078–1083. [142] A. Hambaba, A.E. Huÿ, in: IEEE Aerospace Conference Proceedings, Multi-
[111] Xiaoli Li, X.P. Guan, Time–frequency-analysis-based minor cutting edge resolution error detection on early fatigue cracks in gears, vol. 6, 2000, pp.
fracture detection during end milling, Mechanical Systems and Signal 367–372.
Processing 18 (2004) 1485–1496. [143] Qiao Sun, Ying Tang, Wei Yang Lu, Yuan Ji, Feature extraction with discrete
[112] Hai Qiu, Jay Lee, Jing Lin, Gang Yu, Wavelet filter-based weak signature wavelet transform for drill wear monitoring, Journal of Vibration and
detection method and its application on rolling element bearing prognostics, Control 11 (11) (2005) 1375–1396.
Journal of Sound and Vibration 289 (2006) 1066–1090. [144] S. Loutridis, A. Trochidis, Classification of gear faults using Hoelder
[113] Jing Lin, Liangsheng Qu, Journal of sound and vibration, feature extraction exponents, Mechanical Systems and Signal Processing 18 (2004) 1009–1030.
based on Morlet wavelet and its application for mechanical fault diagnosis, [145] A.N. Robertson, C.R. Farrar, H. Sohn, Singularity detection for structural
Journal of Sound and Vibration 234 (1) (2000) 135–148. health monitoring using holder exponents, Mechanical Systems and Signal
[114] J. Lin, M.J. Zuo, K.R. Fyfe, Mechanical fault detection based on the wavelet Processing 17 (6) (2003) 1163–1184.
de-noising technique, Journal of Vibration and Acoustics 126 (2004) [146] Z. Peng, Y. He, Z. Chen, F. Chu, Identification of the shaft orbit for rotating
9–16. machines using wavelet modulus maxima, Mechanical Systems and Signal
[115] D.B. Percival, A.T. Walden, Wavelet Methods for Time Series Analysis, Processing 16 (2002) 623–635.
Cambridge University Press, Cambridge, 2000. [147] Z.K. Peng, F.L. Chu, Peter, W. Tse, Singularity analysis of the vibration signals
[116] A. Antoniadis, D. Leporini, J.C. Pesquet, Wavelet thresholding for some by means of wavelet modulus maximal method, Mechanical Systems and
classes of non-Gaussian noise, Statistica Neerlandica 56 (4) (2002) 434–453. Signal Processing 21 (2) (2007) 780–794.
ARTICLE IN PRESS
[148] Kenneth C. Chou, Larry P. Heck, A Multiscale Stochastic Modeling Approach [151] W. Hardle, G. Kerkyacharian, D. Picard, A. Tsybakov, Wavelets, Approxima-
to the Monitoring of Mechanical Systems, IEEE Press, New York, 1994. tion and Statistical Applications, Lecture Notes in Statistics, vol. 129,
[149] A.A. Safavi, J. Chen, J.A. Romagnoli, Wavelet-based density estimation and Springer, New York, 1998.
application to process monitoring, Artificial Intelligence in Chemical [152] Brani Vidakovic, Statistical Modeling by Wavelet, Wiley, New York,
Engineering 43 (5) (1997) 1227–1241. 1999.
[150] K. Krishnan Nair, Anne S. Kiremidjian, Time series based structural damage [153] M. Vannucci, Nonparametric density estimation using wavelets: a review.
detection algorithm using Gaussian mixtures modeling, Journal of Dynamic Discussion Paper 95-26, ISDS, Duke University, USA, 1995.
Systems, Measurement, and Control—Transactions of the ASME 129 (May) [154] F. Abramovich, T.C. Bailey, T. Sapatinas, Wavelet analysis and its statistical
(2007) 285–293. applications, The Statistician 49 (2000) 1–29.
View publication stats

Wavelet in TCM Review

Uploaded by

Copyright:

Available Formats

Wavelet in TCM Review

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Wavelet in TCM Review

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

Article in International Journal of Machine Tools and Manufacture · June 2009

Kunpeng Zhu Geok-Soon Hong

SEE PROFILE SEE PROFILE

The user has requested enhancement of the downloaded file.

International Journal of Machine Tools & Manufacture 49 (2009) 537–553

Contents lists available at ScienceDirect

International Journal of Machine Tools & Manufacture

Wavelet analysis of sensor signals for tool condition monitoring:

1. Overview of tool condition monitoring. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537

Nomenclature WTMM wavelet transform modulus maxima

Machining Signals Signal Processing Classifier Tool State

Fig. 1. The framework of TCM.

Approaches References Comments

factor deﬁned, decomposition [69] is a generalization of wavelet decomposition

Q a0 ¼ Central frequency=bandwidth ¼ ðZ=a0 Þ=ðso =a0 Þ ¼ Q a obtained from a given W0 as follows:

Cutting force and it's multiresolution decomposition

0 47Hz 94Hz 188Hz 375Hz 750Hz 1500Hz 3000Hz

Wavelet Wavelet features Tool condition Monitored signal Reference

Wavelet Reference Transform Suitability Comments

CWT [73–75] Linear Stationary, non-stationary Redundant transform, computationally slow

Fig. 6. Wavelet denoising scheme.

wavelet residue probability density

noise probability density

to motor current signal. The wavelet packet features were 25

also studied in [127], while the classiﬁcation is based on a MLP 15

effectiveness of the approach was evaluated under different

Slight wear Severe wear

Holder estimates at time instants T Holder estimates at time instants T

Singularity Spectrum Singularity Spectrum

View publication stats

You might also like