Mechanical Systems and Signal Processing 93 (2017) 80–91

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

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

The monitoring of micro milling tool wear conditions by wear

area estimation
Kunpeng Zhu a,⇑, Xiaolong Yu a,b
a
Institute of Advanced Manufacturing Technology, Hefei Institutes of Physical Science, Chinese Academy of Sciences, Huihong Building, Changwu Middle Road
801, Changzhou 213164, Jiangsu, China
b
Department of Automation, University of Science and Technology of China, Hefei 230026, China

a r t i c l e i n f o a b s t r a c t

Article history: In micro milling, the tool wear condition is key to the geometrical and surface integrity of
Received 22 July 2016 the product. This study proposes a novel tool wear surface area monitoring approach based
Received in revised form 16 January 2017 on the full tool wear image, which can reflect the tool conditions better than the traditional
Accepted 4 February 2017
tool wear width criteria. To meet the challenges of heavy noise, blur boundary, and
Available online 16 February 2017
mis-alignment of the captured tool wear images, this paper develops a region growing
algorithm based on morphological component analysis (MCA) to solve the problems. It
Keywords:
decomposes the original micro milling tool image into target tool images, background
Micro milling
Tool wear area estimation
image and noise image. Then, the region growing algorithm is used to detect the defect
Morphological component analysis and extract the wear region of the target tool image. In addition, rotation invariant features
Region growing are extracted from wear region to overcome the inconsistency of wear image orientation.
The experiment results show that region growing based on MCA algorithm can extract the
wear region of the target tool image effectively and the extracted wear region also has good
indication of tool wear conditions. It also demonstrates that the estimation of wear area
can generalize the tool wear width estimation approach, and yield more accurate results
than the traditional approaches.
Ó 2017 Elsevier Ltd. All rights reserved.

1. Introduction

Micro-milling has wide applications in micro and ultra precision devices for its prominent capabilities in versatile mate-
rial processing and complex 3D surface machining. However, due to the nature of ultra-high speed interrupted cutting and
micro-scale tool, the tool wear appears quickly. The tool wear condition has direct effects on the quality of products. Serious
tool wear can also cause tool breakage, break off and chatter, which would even harm to the machine tool. An effective the
tool wear monitoring is most important [1]. Previous work mainly includes two categories: one of them used indirect
method such as cutting force, vibration and acoustic emission to infer the tool wear condition; another one employed direct
method such as digital image processing techniques in milling tool image to monitor the tool wear condition.
The indirect methods mainly apply sensor signals, including force, vibration, and acoustic emission (AE) to estimate tool
wear conditions. Tansel et al. [2] studied the fracture and wear of vertical milling cutter groove on mild steel and aluminum,
they estimated the wear of the cutting tool by monitoring the feed direction of cutting force on workpiece. They also
controlled the cutting force by dynamically adjusting the cutting parameter to extend the service life of cutting tools. By

⇑ Corresponding author.
E-mail address: kunpengz@hotmail.com (K. Zhu).

http://dx.doi.org/10.1016/j.ymssp.2017.02.004
0888-3270/Ó 2017 Elsevier Ltd. All rights reserved.
 K. Zhu, X. Yu / Mechanical Systems and Signal Processing 93 (2017) 80–91 81

analyzing the time-frequency characteristics of cutting force and dynamic characteristic, Zhu et al. [3,4] proposed methods
based on stochastic process modeling of cutting forces and tool conditions; it achieved good tool wear monitoring results by
considering denoising and discriminant force feature extractions. Others tried vibration analysis. Lamraoui et al. [5] inves-
tigated tool wear monitoring using the vibrations from high speed milling, and showed that the tool wear and the number of
broken teeth on the excitation of structure resonances appeared in its Wigner-Ville representation. Hsieh et al. [6] also
detected the variation of tool wear through vibration analysis, and found the vibration was sensitive to the directions and
the feature selection was important with neural network classification. Another popular tool wear monitoring approach is
by acoustic emission. It has advantage in that the signal measured is a source of engagement where the chip is formed.
Prakash and Kanthababu [7] and Hung and Lu [8] utilized acoustic emission signal in the micro milling tool wear monitoring
field respectively. They claimed that there were strong relationships between the tool wear (flank wear) and acoustic emis-
sion signals. The acoustic emission signal and its frequency spectrum characteristics could monitor the tool wear monitoring
and meet different application aspects. Yen et al. [9] applied AE signals together with intelligent approaches and achieved
satisfactory tool wear classification results. By fusion different sensors, Duro et al. [10] proposed a multi-sensor data fusion
system to enable identification of the best sensor locations for monitoring cutting operations, and it was found this frame-
work could improve signal interpretation and enhance the monitoring system reliability. Malekian et al. [11] developed a
comprehensive system and fused various signal features for the tool wear monitoring in micro milling. The indirect methods
generally has drawbacks with small amplitude (force, vibration), low signal-noise rate (force, vibration), wide band and high
sampling frequency (AE) in the micro milling monitoring. It has always been a difficult task to establish a reliable relation-
ships between the signal features and tool wear states under varying working conditions.
In order to overcome the shortcomings of the indirect methods in cutting tool monitoring, researchers proposed direct
monitoring methods based on tool wear image captured. Due to the development of high speed cameras, this approach
has benefits of high precision, large dynamic ranges and indirect contact, and it attracts more attention in the research.
Kurada and Bradley [12] implemented a direct tool condition monitoring system using two fiber optic lighting and CCD cam-
eras to obtain the tool flank wear images. They put to use a cascade median filter to reduce the image noise and utilized
image segmentation method to extract tool wear area from its background. The error rate was not satisfactory as reported
however.
On the basis of the combination of tool wear measurement and classification, Lanzetta [13] put forward a kind of auto-
matic load sensing system, and the resolution of the sensor reached 0.04 mm/pixel. Zhang et al. [14] developed a technique
for the tool wear measurement based on machine vision to extract the edges of the tool wear region by column scanning. To
achieve a more accurate result, they adopted the sub-pixel edge detection technology to extract the edges. Su et al. [15] stud-
ied the feasibility of the digital image processing method, which was used to measure the micro-drilling tool (the diameter is
0.2 mm) of flank wear. Other researches also tried image analysis approaches such as image reconstruction [16], boundary
detection [17], texture analysis [18,19], and successive image analysis [20]. Kuttolamadom et al. established a volumetric
tool wear measurement method [21] and found that the volumetric tool wear rate was it is closely related to the material
removal rate when milling Ti–6Al–4V [22]. This study devoted to the understanding of the wear evolving in 3D. This system
is rather computationally and physically expensive with 3D surface profiler scanning, and hard to be implemented online/
in-process however. A more convenient way of 3-D measurement was proposed with a phase-shifting method by Wang et al.
[23]. Lachance et al. [24] did a pioneering work to extract and segment the wear surface from grinding wheel image
background their method had solved the precise selection problem of intensity threshold and boundary threshold. However,
their algorithm was did not take account the image morphology, which was inherent in tool wear image and important in
determination of tool wear area.
It is noted that the current literatures mainly measured the tool wear width and regarded it as the tool wear criteria, and
the wear land area has been little investigated in micro-milling to our best knowledge. However, in the micro milling, the
cutting edge has different working load along the edge, and the wear rate varies in different area as a result. At the same
time, the tool wear area is strongly irregular because it is produced in complex mechanical and thermal environment. So,
a more comprehensive geometric properties such as the wear area was needed to serve as an important basis of judging
the degrees of tool wear. To meet this challenges, this paper proposes a novel region growing algorithm based on MCA to
extract micro milling tool wear region and estimate its wear area. It could precisely estimate the tool wear area and handle
the tool wear image that has heavy noise, blur boundary, and uneven grey value distributions. Details of the approach is
developed in the following section.

2. The detection and estimation method of the wear land area of micro milling tool images

Region growing is a kind of image segmentation algorithm and has been widely used in image processing field. It can seg-
ment the connected areas which have same characteristics, as well as provide good boundary information and segmentation
results. In the absence of prior knowledge, the region growing method can get the good performance [25]. However, the tra-
ditional region growing method is not robust when it partitions micro milling tool wear images. If the tool background image
greyscale is uneven in the target region, it will lead to early termination of the regional growth. On the contrary, if the bound-
ary of the image background and the cutting tool are blur, it will give rise to overgrowth of region growing. How to eliminate
the background on the tool wear region is key to the successful defect detection using the region growing algorithm. For the
latest years, the morphological component analysis [26], which synergies with the sparse representation methods [27], has
82 K. Zhu, X. Yu / Mechanical Systems and Signal Processing 93 (2017) 80–91

Micro milling
process

Tool wear
image

MAC Feedback

Threshold
processing
Wear area
estimation
Rotation
invariant

Tool condition

Fig. 1. The micro milling tool condition monitoring by tool wear image processing.

attracted attention for image representation. It can effectively decompose original image into target object image, texture
(background) and noise. In the wear images of micro milling tool, the background generally has similar textures; on the con-
trary, the target tool image is a locally heterogeneous object. In this aspect, the MCA method can be adapted to decompose
original micro milling tool images into various components.
In this paper, by taking with the advantages of the MCA algorithm, a novel region growing algorithm is proposed to detect
and estimate the wear area in micro milling tool images. The original tool wear image is decomposed into the target tool
image, background image and noise image by MCA, an then the region growing method is applied to extract the wear region
and estimate the wear area in the target tool image, which has been separated from original image. Fig. 1 shows the
information flow of micro milling tool condition monitoring using the proposed approach.

2.1. Morphological component analysis (MCA) of the micro milling tool images

The theory of the MCA supposes that the N-sample signal or image is a linear superposition of some morphological
components. In addition, the MCA also assumes that each component can be sparsely represented in an associated basis (also
known as dictionary). Thus, a dictionary can be built by integrating several transforms such that the representation of
current component in corresponding dictionary is sparse but not, or at least not as sparse, in other dictionary. In the other
words, the dictionaries must selected to be mutually incoherent. Thus, the dictionary plays a role of a discriminant between
signal (image) content types, preferring the corresponding component over the other parts. Under this framework, an orig-
inal signal can be represented by the generated over-complete dictionary as:
X
n X
n
S¼ Sk ¼ Dk ak ; k ¼ 1; . . . ; n ð1Þ
k¼1 k¼1

where S is the original signal, Sk the signal component, Dk the over-complete dictionary, and ak the sparse coefficients. This is
an under-determined system and the solution of equation Sk ¼ Dk ak is not unique, as the number of unknowns is greater
than equations. Under the condition of the a following the sparse constraint, a signal MCA model is built to solve above
problem:
X
n
mina1 ;...;an kak kpp
k¼1
  ð2Þ
 Xn 
 
s:t: S  Dk ak  6 s
 k¼1

2

where kak kpp is used to determinate the sparsity of coefficients, and s represents error or noise between original signal and
composite signal.
Given a micro milling tool image I; with MCA the original image could be decomposed as:
I ¼ IO þ IB þ IN ð3Þ
 K. Zhu, X. Yu / Mechanical Systems and Signal Processing 93 (2017) 80–91 83

where IO is the target tool image, IB is background image and IN is noise image. According to the principle of MCA, the target
tool image IO is represented in an over-complete dictionary DO :

aO ¼ arg min kaO k0 s:t: IO ¼ DO aO ð4Þ

where aO is the sparse coefficients of the target tool image. Similarly, the background image IB also has an over-complete
dictionary DB :

aB ¼ arg min kaB k0 s:t: IB ¼ DB aB ð5Þ

where aB is sparse coefficient of background image. Combining the dictionary DO and DB ; an optimization problem is
proposed that:

faopt opt
O ; aB g ¼ arg min kaO k0 þ kaB k0

s:t: kI  DO aO  DB aB k22 6 e ð6Þ

r2e ¼ Var½e < þ1
where e is an acceptable noise level, r2e is noise variance. Fig. 2 represents the MCA decomposition model of micro milling
tool. Formula (6) is a NP-hard problem, and its computational complexity grows exponentially with the column number of
the dictionary. Based on the basis pursuit (BP) algorithm, the l0 norm in equation can be approximated by the l1 norm, and
(6) then turns into a linear programming problem:
2
faopt opt
O ; aB g ¼ arg min kaO k1 þ kaB k1 þ kkI  DO aO  DB aB k2 þ cTVfDO aO g ð7Þ

where k, c are weighting parameters, and TVfDO aO g represents total variation of target image [26].
The choice of the dictionary is the most critical part in the sparse decomposition. In this study, the dictionary is chosen
empirically from known basis based on their respective morphological properties, rather than with dictionary learning algo-
rithms. According to prior knowledge, and without loss of generality, the tool wear image can be assumed to consist of three
parts: target tool, background and noise. It has been investigated in the literatures on how to choose dictionaries to represent
different geometrical (morphological) property images. To represent efficiently isotropic structures, a qualifying choice is the
wavelet transform. The curvelet system is a very good candidate for representing piecewise smooth images [27]. The ridgelet
transform [27] has been shown to be very effective for sparsely representing global lines in an image. For locally oscillating
textures, one can considers the local discrete cosine transform (DCT) [28] and brushlets [27]. These transforms are also
computationally tractable particularly in large-scale applications.
Among them, the local discrete cosine transform (LDCT) [28] is a multiresolution signal analysis method based on over-
lapping orthogonal basis or dual orthogonal basis. This transform divides the signal into N segments by using smooth over-
lapping window function, and then makes cosine transformation on the segments. The LDCT can preserve the low frequency
components and remove most of the high frequency components of the image. It has obtained good results in the image
denoising [28]. The DCT transform of tool image I(x,y) is formulated as:
   
XX
M1 N1
ð2x þ 1Þup ð2y þ 1Þv p
Cðu; v Þ ¼ aðuÞaðv Þ Iðx; yÞ  cos cos ð8Þ
x¼0 y¼0
2M 2N

where Cðu; v Þ is the DCT coefficient, with u ¼ 0; 1; 2 . . . ; M  1, v ¼ 0; 1; 2 . . . ; N  1. On the other hand, the curvelet transform
[29] has been studied for image representation. The Curvelet transform is highly anisotropic, and has the ability to express

Decomposition

Over-complete
dictionary
Target
image

Background
image Target Background
Noise dictionary dictionary
image

Decomposition

Fig. 2. The MCA decomposition model of micro milling tool.

84 K. Zhu, X. Yu / Mechanical Systems and Signal Processing 93 (2017) 80–91

the image edge information preferably against the traditional two dimensional wavelet transform. The Curvelet transform of
tool image I(x,y) is [29]:
Z
Cðj; l; kÞ ¼ hIx;y ; uj;l;k i ¼ Ix;y uj;l;k dxdy ð9Þ
R2

where Cðj; l; kÞ is the curvelet transform coefficient, and uj;l;k is wavelet basis.
In the wear images of micro milling tool, it has been observed that the background generally has self-similar texture with
oscillating patterns, while the target tool image is locally heterogeneous with piecewise smooth structure. So, the local dis-
crete cosine basis is chosen to form the background image dictionary, and the curvelet basis selected for the target tool image
dictionary.

2.2. Region growing algorithm based on MCA

Region growing is a region-based image segmentation method. It achieves image segmentation by grouping neighboring
pixels as a region with some measurements. In region growing, a pixel or a small group of pixels are picked as the seed. The
regions are then grown from these seed points to their neighboring points based on a region similarity criterion, such as pixel
intensity, grey scale texture, or color etc. This process is iteratively computed until all pixels belong to some region.
In this paper, a novel region growing method based on MCA is proposed to extract the wear region of micro milling tool
images and estimate wear area. Without loss of generality, the approach assumes that the original image consists of target
tool image, background and noise. By applying the MCA, it extracts the target tool image and then applies region growing
method on it. In comparison to the traditional region growing algorithm, the side effect of noise and background for segmen-
tation in original images is eliminated. Instead of solving the optimization problem in (7), define IO ¼ DO aO and IB ¼ DB aB ,
1
assuming Dþ ¼ DT ðDDT Þ the Moore-Penrose pseudo-inverse of D, so aO , aB can be recovered as aO ¼ Dþ þ
O IO , aB ¼ DB IB respec-
tively. Put this back into (7) can obtain:
þ þ 2
fDopt opt
O ; DB g ¼ arg min kDO IO k1 þ kDB IB k1 þ kkI  IO  IB k2 þ cTVfIO g ð10Þ
^ O from aO by iteration threshold d, and reconstruct target tool image:
Select a

IO ðx; yÞ ¼ Dopt ^
O ðx; yÞaO ðx; yÞ until d < k ð11Þ

where k is termination condition threshold, and the d is updated by mean of max (MOM) method:

1
dt ¼ ½kDþB ðI  ItO Þk1 þ kDþO ðI  ItB Þk1  ð12Þ
2
Then, the region grow method is applied to target tool image IO ðx; yÞ:

1 jIO ðx; yÞ  Sðx; yÞj < T
g Q ðx; yÞ ¼ ð13Þ
0 jIO ðx; yÞ  Sðx; yÞj P T

where Sðx; yÞ is the seed array, and g Q is the segmented image. Finally the segmented image hðx; yÞ is formed by merging each
seed point with all the 1-valued points that have 8-connected neighbors.
The foremost point is to determine the seed points, as too few or too many seeds will result in under- or over-segmented
images. In the actual image of the cutting tool, it is easy to observe that the defect region is brighter than other parts. So
thresholding is applied to extract the seed points, by using a threshold set at a high percentile in the target tool image.
The next step is to specify an attribute Q at location ðx; yÞ. This study applies the grey scale difference as a measure of
similarity. It is determined to be Q ¼ TRUE if the absolute difference of the grey scale difference between the seed and pixel
at ðx; yÞ is 6 T, otherwise, Q ¼ FALSE, where T is a specified threshold. The algorithm is shown in Table 1.

Table 1
Algorithm of Region growing algorithm based on MCA.

Input: Original image Iðx; yÞ, target dictionary DO , background dictionary DB , weight k, a seed array Sðx; yÞ, an attribute Q .
Initialization:
 Initial residual R ¼ 0.
 Initial seed array Sðx; yÞ ¼ 0; 8x; y.
Main steps:
^B.
 Update IB ðx; yÞ by fixing IO ðx; yÞ, compute residual R ¼ Iðx; yÞ IO ðx; yÞ, and reconstruct IB ðx; yÞ ¼ DB a
^O .
 Update IO ðx; yÞ by fixing IB ðx; yÞ, compute residual R ¼ Iðx; yÞ IB ðx; yÞ, and reconstruct IO ðx; yÞ ¼ DO a
 Label pixels of connected components in Sðx; yÞ as 1, and other 0.
 Form an image g Q ðx; yÞ with g Q ðx; yÞ ¼ 1 if the input image satisfies the given attribute Q ; otherwise, let g Q ðx; yÞ ¼ 0.
 Get the segmented image hðx; yÞ by merging each seed point with nearest 8-neighours.
Output: The segmented image hðx; yÞ.
 K. Zhu, X. Yu / Mechanical Systems and Signal Processing 93 (2017) 80–91 85

90°
0° 90° 180° 270° 0°

180° 0° ln

270° Log-Polar coordinates

Cartesian coordinates

Fig. 3. The Log-Polar transform diagram.

2.3. Rotation invariant feature extraction of wear area

In the tool condition monitoring, the tool keeps rotating at high speed and the tool wear images are captured in different
angular position. As a result, it’s important to extract a feature that achieves rotation invariance for the tool wear region esti-
mation. In this paper, the combination of Fourier transform and Log-Polar transform (LPT) is applied to reach rotation invari-
ant image analysis [30]. The LPT solves it by mapping the image pixel (x; y) in Cartesian coordinate to Log-Polar system (q, h),
namely:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
q ¼ ln ðx  x0 Þ2 þ ðy  y0 Þ2 ð14Þ
 
y  y0
h ¼ arctan ð15Þ
x  x0
As shown in Fig. 3, the LPT transforms the rotation information in Cartesian coordinates into translation information in
Log-Polar coordinates. It can eliminate the directional sensitivity of the Fourier transform to realize rotation invariance.

3. Experimental results and discussions

In this study, there were 24 milling cutter images captured and analyzed, which had different wear degrees under varied
cutting conditions with a high precision milling center. The diameter of micro milling tool is 0.8 mm. The tool wear images
were captured under random angular position with a high resolution optical device. All the image resolution was set
1024  768Þ and, Fig. 4(a) shows the experiment setup and Fig. 4(b) is the measurement setup. Fig. 4(c) shows the bottom
view of the micro milling tool under different wear conditions. The wear in the minor cutting edge is significant in micro-
milling and easy to be monitored and implemented with optical devices. In most studies, the final wear was taken as the
average of width of both ordinates [2,3], while the area (wear land) is estimated and considered as the evaluation criteria
in this study. Similar to the tool wear phenomena in the conventional machining, the minor cutting edges and wear area
also have three stages in the wear progression, which are to be discussed in details in the later sections. The experiment
results include three parts: (1) the decomposition results of micro milling tools by using MCA; (2) the comparison of the
results between region growing algorithm based on MCA and traditional region growing algorithm; (3) the estimation of
different wear area of micro milling tools.

3.1. The decomposition of tool image by MCA

The MCA is a coarse-to-fine iterative process. In practice, given an image I, its forward transform will be computed by
multiplying it by D ¼ ðDO ; DB Þ. In addition, the tool wear image can be reconstructed from its coefficients a ¼ ðaO ; aB Þ. In fact,
the computation of the matrix D is implemented as fast implicit analysis and synthesis operators by taking an image I and
returning DT I ¼ TI, or by taking a coefficient vector a and returning Da. In the case D consists of only orthonormal basis, the
inverse of the analysis operator is trivially T1 ¼ D; on the other side, if D is a tight frame, then the frame operator satisfies
DDT ¼ cI, where c > 0 is the tight frame constant. In this condition, the D equals to TT and is the Moore-Penrose pseudoin-
verse transform. It turns out that Tþ a is the reconstruction operation implemented in the experiment. In the algorithm, a
threshold d is initialized firstly. The decomposition will be terminated by stopping iteration when d > k. In this paper, the
k is set as kre , where re is the noise standard deviation in (6), and k ¼ 3. The target image IO is slightly modified by applying
the total variation by IO ¼ IO  lcð@TVfIO g=@IO Þ, where the parameter l is chosen by the line-search algorithm that mini-
mizes the overall penalty function [31] so as to remove the correlated background noise.
Empirical studied have been carried out for the parameter setting. The window size of the LDCT and the parameter of
Curvelet wavelet are finally set as 16 and 6, respectively, and iterative steps set at 100. The decomposition result is shown
in Fig. 5. Fig. 5(a) represents three original tool images with different rotation angles. Fig. 5(b) shows the target tool image
after applying the MCA, It can be observed from Fig. 5 that the target image has less noise than the original image, and the
white border of cutting tools has been greatly weakened, which is useful for the further image segmentation. Fig. 5(b) also
86 K. Zhu, X. Yu / Mechanical Systems and Signal Processing 93 (2017) 80–91

Microscope

Tool
Micro Wear
Spindle
tool
System Tool
State
Holder

Digital
Camera
Tool
workpiece Digital Computer
Camera

(a) the experiment setup

800µm

120µm
196µm
59µm
73µm

Wear Area Wear Area

The bottom view

(b) the optical measurement (c) bottom view of micro milling tool
Fig. 4. The experiment setup and tool wear measurements.

(a) original image (b) target tool (c) background (d) noise

Fig. 5. The decomposition results of micro milling tool images via MCA.
 K. Zhu, X. Yu / Mechanical Systems and Signal Processing 93 (2017) 80–91 87

Table 2
The statistics of noise.

Statistics Mean Variance Skewness Kurtosis

Noise 1 131.5 8.53 0.0024 3.0016
Noise 2 126.9 9.89 0.0015 2.9939
Noise 3 149.6 7.93 0.0042 2.9961

5
x 10
2
noise
Gaussian fit
1.5
pixel number

0.5

0
0 50 100 150 200 250
gray level

Fig. 6. Noise grey value histogram.

show that this approach is invariant to the orientation of the image, and can handle the difficulties of estimation that the
wear images are captured at random angular position.
Fig. 5(c) is the extracted background image. It is observed that the background is slightly correlated to the original image.
The regularization parameter c determines background noise, and a proper choice of it as above would minimize the effect of
background. The regularization parameter k controls the (random) noise level. Fig. 5(d) is the noise image, and its statistics
are shown in Table 2 and in Fig. 6. The results of Gaussian fitting of first noise image indicate that they basically fulfil the
Gaussian distribution. Table 2 has enumerated the mean, variance, skewness and kurtosis of above three noise image. To this
end, the MCA method can eliminate the background and noise effectively from the original tool wear images.

3.2. The segmentation of tool wear image

The extracted target tool image (as Fig. 5(b)) are then segmented by the region growing algorithm developed in
Section 2.2. The key of the region growing algorithm is to find initial seed point (namely pixel grey value). Ideally, the
minimum grey value m of wear region and the maximum grey value M of non-wear region is known in advance, then
(m þ M)/2 can be regarded as the initial seed point, and if the selected pixel grey value is larger than that value, the segmen-
tation algorithm will wrongly take many of non-wear area as tool wear part, and vice versa. However, the two values are

x 104

1.5

0.5

0 50 100 150 200 250

Fig. 7. The histogram of target tool image.

88 K. Zhu, X. Yu / Mechanical Systems and Signal Processing 93 (2017) 80–91

(a) original image (b) traditional algorithm (c) our algorithm

Fig. 8. The segmentation results of region growing algorithm based on MCA and traditional region growing algorithm.

unknown in practice. In this study, it is known a priori that some pixels may be have largest value (255) in the target tool
image. Based on above information, the initial seed array S is set as 1, where the value 255 is mapped to 1 with normaliza-
tion. As shown in Fig. 7, which is the histogram of the first image in Fig. 8, the threshold value is selected as 55, which rep-
resents the difference between the first major peak valley (grey value is 200) and maximal grey value 255. The first major
peak valley represents the highest grey value in the dark background region. The rest of extracted tool images are handled

(a) binary image (b) rotation invariant (c) Fourier power spectrum
Fig. 9. The Log-Polar transform and Fourier power spectrum.
 K. Zhu, X. Yu / Mechanical Systems and Signal Processing 93 (2017) 80–91 89

similarly. Fig. 8 shows the comparison of segmentation results between region growing algorithm based on MCA and tradi-
tional region growing algorithm [25].
In the wear images of micro milling tool, the grey level of tool boundary and white speck noise is similar to tool wear
region. Therefore, the traditional region growing method may get wrong segmentation results due neglecting the influence
of background and noise. It may treat the tool boundary and noise as a part of the tool wear defect. Fig. 8(b) is the result
based on traditional region growing algorithm, it’s obvious that the wear region is over-segmented. The region growing algo-
rithm based on MCA can solve the above problems effectively. As shown in Fig. 8(c), it reduces the influence of background
and noise on tool images. The correctness of the segmentation is greatly improved.
Fig. 9 shows the results of Log-Polar transform, which acts on the binary image of tool wear region. The Fourier power
spectrum is extracted to be rotation invariant. From Fig. 9(a) and (b), it is observed that the rotation information has been
transformed into translational information in tool images. Additionally, the Fourier power spectrum of the images (b) are
shown in Fig. 9(c). The top and bottom figures of Fig. 9(a) are essentially the same. According to the power spectrum’s
translational invariance property, it verifies that the two images are the same in Fig. 9(b).

3.3. The estimation of the wear area

It’s generally observed that the more severe tool wear corresponding to greater wear area. The key to this problem is how
to calculate the area. Given an image, its resolution is pre-determined and the number of pixels can reflect the size of image.
Therefore, it only needs to extract the number of wear zone pixels, which represents wear area. After the image binarizing, it
is known from the experiment results that the grey value of wear region is 255 (white), and other parts are 0 (black). In this
aspect, the proportion of the number of wear region pixels in the full image can be applied to estimate the wear area, that is:
n
a¼A ð16Þ
N
where a is the estimation of wear area, A is actual area of whole image, n is the numbers of pixels of wear area, and N is
the total number of the full image pixels. Table 3 shows the estimation of wear area of tool image under different wear
degrees. It shows that the wear zone is in triangular shape, and with the progressing of tool wear, the triangular area expands
gradually. This is mainly due to the friction and ploughing effect between the cutting tool and workpiece. The second column
is the segmented images after applying the MCA and region growing approach. It is observed that there existed background
noise due adhered chips. These chips are estimated as wear areas. While in general, these areas are randomly and sparsely
distributed in the segmented images. Their values are rather low proportion to the real wear area, and could be neglected.

3.4. Discussions

Fig. 10 shows the tool wear progression with the time. It has been found that the wear area estimation (square line) is
generally consistent in the trend with the manual width measurements (star line). The left ordinate is wear width which

Table 3
The estimation of wear area of tool image with different wear degrees.

Original image Wear image The number of pixels in wear region The wear area ratio (%) The estimation of wear area (lm2)
774 0.35 1486.6

3055 1.37 5818.9

4019 1.80 7645.3

6356 2.85 12105.1

9696 4.34 18433.7

90 K. Zhu, X. Yu / Mechanical Systems and Signal Processing 93 (2017) 80–91

4
x 10
100 2
wear width (estimations)
90 1.8
wear width (measurements)
80 wear area (estimations) 1.6

70 1.4

Wear area (µm2)

Wear width (µm)
60 1.2

50 1

40 0.8

30 0.6

20 0.4

10 0.2

0 0
1 2 3 4 5 6 7 8 9 10
Time (min)

Fig. 10. Tool wear width and area progression with time.

represents the maximum number of the wear pixel along the longitude axis, while the right ordinate represents wear area in
Fig. 4c. It is noted that the two ordinates are represented in different scales. Aparently, the wear width and wear area
increases gradually with time progressing.
In an extreme condition, when only piexles along the ordinate of the wear area (along the longitude axis in Fig. 4c) are
counted (circle line), it converges to the tool wear width estimation. These width estimations are obtained by estimating the
numbers of white pixels of each column in the binarized image. The maximum value and second-largest value represent
the maximum wear width on both sides of the cutting tool respectively. The average value of these two is the wear width.
The two figures shows a good conincidence. The experimental results indicate that the proposed wear area algorithm can
generalize the tool wear width method effectively. In Fig. 10, it also shows that the slope of the wear area estimation is
higher than that of wear width in the accelerated wear region, which means that the wear area increases proportionally
faster than the wear width, and measures more accurate the tool wear state variations.
The wear area is more accurate than the width to evaluate the tool wear degrees as it includes more tool wear informa-
tion. In the micro milling, the cutting edge has different working load along the edge (edge slice), and the wear rate varies in
different area as a result. The cutting tip has highest stress but with least heat diffusion, the wear is most significant in this
area. While for the cutting edge close to the exterior surface of the workpiece, the stress is lighter and heat diffuses faster, the
wear land is much less in this zone. As a result the wear area estimation considers the variations of all the wear zone and
reflects better the real conditions of tool wear.

4. Conclusion and future work

In this paper, a novel region growing method based on morphological component analysis is proposed to detect the wear
region and estimate the wear area in micro milling tool. Compared to the other image processing based approaches, the
proposed technique greatly reduces the side effects of image background and noise for tool condition monitoring. The exper-
imental results have validated the effectiveness of this approach. It also shows that this approach generalizes the tool wear
width approach, and is more accurate than the width estimation as it counts the variations of all the wear zones and reflects
real tool wear conditions. In the future work, a criteria will be formulated to quantize the wear degrees according to the wear
area estimated. In addition, a multiscale MCA algorithm would be studied for the sub-pixel resolution enhancement.

Acknowledgment

This project is supported by the CAS100 Talents Program, the Chinese Academy of Sciences (CAS), and the National Nat-
ural Science Foundation of China (Grant No. 51475443).

References

[1] R. Teti, K. Jemielniak, G. O’Donnell, D. Dornfeld, Advanced monitoring of machining operations, CIRP Ann. Manuf. Technol. 59 (2) (2010) 717–739.
[2] W.Y. Bao, I.N. Tansel, Modeling micro-end-milling operations. Part III: influence of tool wear, In. J. Mach. Tools Manuf. 40 (15) (2000) 2193–2211.
[3] K.P. Zhu, Y.S. Wong, G.S. Hong, Multi-category micro-milling tool wear monitoring with continuous hidden Markov models, Mech. Syst. Signal Process.
23 (2) (2009) 547–560.
 K. Zhu, X. Yu / Mechanical Systems and Signal Processing 93 (2017) 80–91 91

[4] K.P. Zhu, B. Vogel-Heuser, Sparse representation and its applications in micro-milling condition monitoring: noise separation and tool condition
monitoring, Int. J. Adv. Manuf. Technol. 70 (1–4) (2014) 185–199.
[5] M. Lamraoui, M. Thomas, M. El Badaoui, Cyclostationarity approach for monitoring chatter and tool wear in high speed milling, Mech. Syst. Signal
Process. 44 (1–2) (2014) 177–198.
[6] W.H. Hsieh, M.C. Lu, S.J. Chiou, Application of back-propagation neural network for spindle vibration-based tool wear monitoring in micro-milling, Int.
J. Adv. Manuf. Technol. 61 (1–4) (2012) 53–61.
[7] M. Prakash, M. Kanthababu, In-process tool condition monitoring using acoustic emission sensor in microendmilling, Mach. Sci. Technol. 17 (2) (2013)
209–227.
[8] C.W. Hung, M.C. Lu, Model development for tool wear effect on AE signal generation in micromilling, Int. J. Adv. Manuf. Technol. 66 (9–12) (2013)
1845–1858.
[9] Chia-Liang Yen, Ming-Chyuan Lu, Jau-Liang Chen, Applying the self-organization feature map (SOM) algorithm to AE-based tool wear monitoring in
micro-cutting, Mech. Syst. Signal Process. 34 (1–2) (2013) 353–366.
[10] J.A. Duro, J.A. Padget, C.R. Bowen, H.A. Kim, A. Nassehi, Multi-sensor data fusion framework for CNC machining monitoring, Mech. Syst. Signal Process.
66–67 (1) (2016) 505–520.
[11] M. Malekian, S.S. Park, M.B.G. Jun, Tool wear monitoring of micro-milling operations, J. Mater. Process. Technol. 209 (10) (2009) 4903–4914.
[12] S. Kurada, C. Bradley, A machine vision system for tool wear assessment, Tribol. Int. 30 (4) (1997) 295–304.
[13] M. Lanzetta, A new flexible high-resolution vision sensor for tool condition monitoring, J. Mater. Process. Technol. 119 (1) (2001) 73–82.
[14] J. Zhang, C. Zhang, S. Guo, L. Zhou, Research on tool wear detection based on machine vision in end milling process, Prod. Eng. 6 (4–5) (2012) 431–437.
[15] J.C. Su, C.K. Huang, Y.S. Tarng, An automated flank wear measurement of microdrills using machine vision, J. Mater. Process. Technol. 180 (1–3) (2006)
328–335.
[16] M. Szydlowski, B. Powalka, M. Matuszak, P. Kochmański, Machine vision micro milling tool wear inspection by image reconstruction and light
reflectance, Precision Eng. 44 (1) (2016) 236–244.
[17] M. Sortino, Application of statistical filtering for optical detection of tool wear, Int. J. Mach. Tools Manuf. 43 (5) (2003) 493–497.
[18] D. Kerr, J. Pengilley, R. Garwood, Assessment and visualisation of machine tool wear using computer vision, Int. J. Adv. Manuf. Technol. 28 (7) (2006)
781–791.
[19] S. Dutta, A. Datta, N.D. Chakladar, S.K. Pal, S. Mukhopadhyay, R. Sen, Detection of tool condition from the turned surface images using an accurate grey
level co-occurrence technique, Precision Eng. 36 (3) (2012) 458–466.
[20] W. Wang, Y.S. Wong, G.S. Hong, Flank wear measurement by successive image analysis, Comput. Ind. 56 (8–9) (2005) 816–830.
[21] M.A. Kuttolamadom, M.L. Mears, T.R. Kurfess, On the volumetric assessment of tool wear in machining inserts with complex geometries – Part 1: Need,
methodology, and standardization, ASME J. Manuf. Sci. Eng. 134 (5) (2012) 0510021-8.
[22] M.A. Kuttolamadom, P. Mehta, M.L. Mears, T.R. Kurfess, Correlation of the volumetric tool wear rate of carbide milling inserts with the material
removal rate of Ti-6Al-4V, ASME J. Manuf. Sci. Eng. 137 (2) (2015) 0210211-8.
[23] W.H. Wang, Y.S. Wong, G.S. Hong, 3-D measurement of crater wear by phase-shifting method, Wear 261 (2) (2006) 164–171.
[24] S. Lachance, R. Bauer, A. Warkentin, Application of region growing method to evaluate the surface condition of grinding wheels, Int. J. Mach. Tools
Manuf. 44 (7–8) (2004) 823–829.
[25] S.A. Hojjatoleslami, J. Kittler, Region growing: a new approach, IEEE Trans. Image Process. 7 (7) (1998) 1079–1084.
[26] J.L. Starck, M. Elad, D.L. Donoho, Image decomposition via the combination of sparse representations and a variational approach, IEEE Trans. Image
Process. 14 (10) (2005) 1570–1582.
[27] J.L. Starck, F. Murtagh, J.M. Fadili, Sparse Image and Signal Processing: Wavelets, Curvelets, Morphological Diversity, Cambridge University Press, 2010.
[28] E. Wesfreid, M.V. Wickerhauser, Adapted local trigonometric transforms and speech processing, IEEE Trans. Signal Process. 41 (12) (1993) 3596–3600.
[29] E.J. Candes, L. Demanet, D.L. Donoho, L. Ying, Fast discrete curvelet transforms, SIAM Multiscale Model. Simul. 3 (5) (2003) 861–899.
[30] Q. Wu, S. Wang, X. Zhang, Log-polar based scheme for revealing duplicated regions in digital images, IEEE Signal Process. Lett. 18 (10) (2011) 559–562.
[31] Y.M. Huang, M.K. Ng, Y.W. Wen, A new total variation method for multiplicative noise removal, SIAM J. Imag. Sci. 2 (1) (2009) 20–40.