Journal of Manufacturing Processes 48 (2019) 66–76

Contents lists available at ScienceDirect

Journal of Manufacturing Processes

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

Data-driven smart manufacturing: Tool wear monitoring with audio signals T

and machine learning
Zhixiong Lia, Rui Liub, Dazhong Wuc,*
a
Department of Mechanical and Aerospace Engineering, University of Central Florida, Orlando FL 32816, USA
b
Department of Mechanical Engineering, Rochester Institute of Technology, Rochester, NY 14623, USA
c
Department of Mechanical and Aerospace Engineering, Department of Industrial Engineering and Management Systems, University of Central Florida, Orlando FL 32816,
USA

A R T I C LE I N FO A B S T R A C T

Keywords: Tool wear in machining could result in poor surface finish, excessive vibration and energy consumption.
Tool wear monitoring Monitoring tool wear in real-time is crucial to improve manufacturing productivity and quality. While numerous
Smart manufacturing sensor-based tool wear monitoring techniques have been demonstrated in laboratory environments, few tool
Audio signal processing wear monitoring systems have been deployed in factories because it is not realistic to install some of the im-
Machine learning
portant sensors such as dynamometers on manufacturing machines. To address this issue, a novel audio signal
Noise reduction
processing approach is introduced. This technique does not require expensive sensors but audio sensors only. A
blind source separation method is used to separate source signals from noise. An extended principal component
analysis is used for dimensionality reduction. Real-time multi-channel audio signals are collected during a set of
milling tests under varying cutting conditions. The experimental data are used to develop and validate a pre-
dictive model. Experimental results have shown that the predictive model is capable of classifying tool wear
conditions with high accuracy.

1. Introduction certain condition monitoring signals such as vibration, acoustic emis-

sion, cutting force, spindle power, and temperature signals [7]. Current
Cutting tools are used to remove material from workpieces by tool wear indirect monitoring techniques fall into two categories:
means of shear deformation in various machining processes such as model-based and data-driven methods [8]. Model-based methods are
turning, milling, and drilling. According to the U.S. Cutting Tool generally effective for predicting tool wear in the cases where in-depth
Institute, manufacturers in the U.S. consumed over $200 million worth knowledge of wear mechanisms are available. However, in practice tool
of cutting tools during April 2018. Excessive tool wear could result in wear mechanisms vary depending on cutting conditions, which makes
substantial decreases in dimensional accuracy, significant increases in effective tool wear prediction very difficult. In addition, certain statis-
energy consumption, and eventually total breakage of cutting tools due tical distributions must be assumed for model-based methods to de-
to excessive cutting forces and vibrations, intensive stresses and tem- velop close-form analytical solutions. These assumptions usually cannot
perature, as well as massive fracture at cutting edges. Health mon- be justified. To complement model-based methods, data-driven
itoring and predictive analytics techniques are crucial to monitoring the methods predict tool wear using predictive models trained by machine
health conditions of cutting tools as well as predicting tool wear [1–4]. learning or pattern recognition algorithms. In addition, existing sensor-
Current tool wear measurement techniques can be classified into based tool wear monitoring techniques have a primary limitation.
two categories [5]: direct and indirect methods. With the direct While various sensor-based condition monitoring signals have been
methods, tool wear is measured directly using optical or radioactive demonstrated to be effective in monitoring and predicting tool wear in
techniques [6]. The direct methods have two primary limitations: (1) it laboratory environments, it is not realistic for manufacturers to retrofit
is not realistic to measure tool wear using imaging systems during CNC machines with expensive sensors such as dynamometers.
machining operations in real-world factory environments; (2) mea- To address these issues, a novel signal processing technique is in-
suring tool wear using imaging systems is very time consuming and troduced to denoise audio sensor-based signals collected during milling
expensive. With the indirect methods, tool wear is estimated based on operations. A new feature extraction method is introduced based on

⁎
Corresponding author.
E-mail addresses: [email protected] (Z. Li), [email protected] (R. Liu), [email protected] (D. Wu).

https://doi.org/10.1016/j.jmapro.2019.10.020
Received 5 October 2018; Received in revised form 18 July 2019; Accepted 14 October 2019
1526-6125/ © 2019 The Society of Manufacturing Engineers. Published by Elsevier Ltd. All rights reserved.
Z. Li, et al. Journal of Manufacturing Processes 48 (2019) 66–76

time-frequency domain analysis and adaptive kernel principal compo- obtain time-frequency-energy distribution of the signals. Experimental
nent analysis (AKPCA) to process the denoised audio signals. The re- results have shown that WPT and HHT are capable of detecting chatter
mainder of this paper is organized as follows: Section 2 presents related effectively. However, FFT, HHT, and WPT are not able to process
work on audio-based tool wear monitoring systems. Section 3 in- multiple channel signals. To address this issue, Zhu et al. [21] develop a
troduces the blind source separation, feature extraction and selection tool wear monitoring system for micro-milling using ICA. ICA is capable
methods. Section 4 presents an experimental setup and the design of of separating a multivariate signal into additive components by as-
milling tests. Section 5 presents experimental results. Section 6 provides suming that the subcomponents are non-Gaussian signals and these
conclusions and future work. signals are statistically independent from each other. Experimental re-
sults have shown that ICA outperforms traditional denoising methods
2. Related work such as wavelets. Shi et al. [22] developed a sound signal processing
technique for tool breakage detection in face milling processes based on
2.1. Tool wear monitoring using audio signals empirical mode decomposition (EMD) and ICA. Experimental results
have shown that the proposed method is capable of detecting tool
Weller et al. [9] developed an electronic-mechanical system that breakage by denoising sound signals during face milling. One of the
uses sonic signals to detect the degree of cutting edge wear in turning limitations of ICA is that it assumes that subcomponents are in-
operations. Experimental results have shown that the system can detect dependent from each other, which may not hold true [23,24]. To ad-
tool wear when cutting AISI 1045 steel with cutting tools made from dress this issue, bounded component analysis (BCA) has recently been
cemented carbide. Mannan et al. [10] developed a sound analysis-based proposed to denoise multivariate signals [23–25].
technique to monitor the conditions of cutting tools. The proposed While previous studies have attempted to develop audio signal-
technique can process audio signals generated during machining pro- based tool wear monitoring techniques, little research has been re-
cesses as well as identify the correlation between tool wear and sound ported on monitoring tool wear in milling using multivariate audio
patterns. Experimental results have shown that cutting tools with sharp, signals. In addition, no studies have been conducted to improve tool
semi-dull, and dull conditions can be detected using this technique. wear monitoring by denoising multivariate audio signals. To fill the
Delio et al. [11] introduced an approach that can detect chatter using research gap, a novel audio-based tool wear monitoring approach is
audio signals. Experimental results have shown that the performance of introduced. This method is capable of denoise multivariate audio sig-
the audio signal-based chatter detection approach is comparable with nals as well as classify tool wear conditions using data-driven predictive
other methods based on dynamometers, accelerometers, and displace- models.
ment probes. Salgado and Alonso [12] developed a tool condition
monitoring system for online tool wear monitoring in turning processes.
Singular spectrum analysis was used to extract features from feed motor 3. Signal processing and machine learning
current and sound signals collected during turning. A support vector
machine-based method was used to estimate tool wear by analyzing the 3.1. Computational framework
extracted features. Aliustaoglu et al. [13] developed a tool wear mon-
itoring technique based on two-stage fuzzy logic and audio signals Fig. 1 illustrates a computational framework of the proposed audio
collected via microphones. A set of drilling experiments was conducted signal processing approach for tool wear monitoring. First, the raw
on a four-axis CNC machining center. Audio signals were collected multi-channel audio signals are decomposed into multiple wavelet sub-
through a microphone. Experimental results have shown that the two-
stage fuzzy logic is capable of identifying tool wear conditions. Ub-
hayaratne et al. [14] developed an approach to tool wear monitoring in
sheet metal stamping using an audio signal processing technique. A
semi-blind signal extraction technique was used to preprocess and de-
noise the audio signals collected from stamping operations. Seemuang
et al. [15] developed a tool condition monitoring system that predicts
tool wear by measuring audio signals generated by machine spindles. A
low-cost microphone was used to collect the audio signals in drilling.
Kothuru et al. [16] developed a tool wear monitoring approach that can
classify tool wear conditions using support vector machines (SVM).
Audio signals collected during milling processes were transformed to
features in the frequency domain. Experimental results have shown that
the prediction accuracy is above 90%.

2.2. Noise reduction in tool wear monitoring

According to a literature review [17,18], the most popular signal

processing techniques in machining include fast Fourier transform
(FFT), Wavelet transform, Hilbert-Huang transform (HHT), principal
component analysis (PCA), and independent component analysis (ICA).
Fang et al. [19] investigated the effects of tool edge wear on cutting
forces and vibrations in high-speed finish machining using wavelet
transform analysis. A set of turning experiments was performed to
collect vibration signals. Both FFT and wavelet transform were used to
denoise the vibration signals. Experimental results have shown that
wavelet transform is more effective than FFT. Cao et al. [20] developed
a chatter detection method for end milling processes using wavelet
package transform (WPT) and HHT. WPT was used to denoise the Fig. 1. A computational framework for the proposed audio signal processing
measured signals. The denoised signals were analyzed using HHT to approach.

67
Z. Li, et al. Journal of Manufacturing Processes 48 (2019) 66–76

bands using WPT. Second, an extended convolutive bounded compo- 3.3. Multivariate synchrosqueezing transform
nent analysis (ECBCA) is used to separate source signals from the wa-
velet sub-band signals. Third, the separated source signals are decom- The source audio signals separated by ECBCA from the mixed sig-
posed into time-varying oscillatory components using a multivariate nals can be modeled as a sum of components with a varying amplitude
synchrosqueezing transform (MSST) in order to denoise. Fourth, eight and instantaneous frequency. Therefore, the retrieval of the compo-
(8) statistical features (i.e., max, min, mean, median, moment, skew- nents of a multicomponent signal is a critical issue in audio signal
ness, kurtosis, and standard deviation) are extracted in the time-fre- processing. Time-frequency analysis has been used to process non-sta-
quency domain. Another feature is the sum of frequency-amplitude in a tionary signals in both the time and frequency domains simultaneously.
certain frequency band that contains the significant frequency harmo- While wavelet transform and short-time Fourier transform (STFT) are
nics. Thus, nine features are extracted from the multi-channel audio the most popular time-frequency analysis methods, these methods are
signal. Fifth, an adaptive kernel principal component analysis (AKPCA) not effective in processing signals with oscillatory properties [27]. To
method is used to decompose the extracted features into a set of linearly address the limitation of wavelet transform and STFT, the multivariate
uncorrelated components. The first two components out of the linearly synchrosqueezing transform (MSST) is used to characterize multivariate
uncorrelated components are fed into a classifier in order to classify signals with time-varying oscillatory properties [27,28]. The MSST
tool wear conditions. method is briefly introduced as follows:
For a signal x(t), its wavelet-based synchrosqueezing is expressed as
3
3.2. Extended convolutive bounded component analysis Sw (ω, b) = ∫ C (a, b) a− 2 δ (ωx (a, b) − ω) da (4)

As shown in Fig. 1, one of the important steps of the computational Where, Sw (ω, b) are the synchrosqueezing transform (SST) coefficients,
framework is blind source separation (BSS). BSS refers to the techniques C (a, b) are the wavelet coefficients, a is a scale factor, b is a shift factor,
that separate a set of source signals from a set of mixed signals. The ω is the frequency, and ωx is the instantaneous frequency.
most popular BSS method is ICA [23,24] that can separate independent The SFTF-based synchrosqueezing can be described by
sources. Bounded component analysis (BCA) is an extension of ICA. 1
BCA is capable of separating both independent and dependent sources. Ss (ω, τ ) =
ω (0)
∫ ST (τ , η) δ (ω − ωx (τ , η)) dη (5)
The convolutive bounded component analysis (CBCA) was introduced
to process signals that are space-time mixtures of the original sources Where, Ss (ω, τ ) are the SST coefficients, ST (τ , η) is the STFT compo-
[23,24]. In this paper, an ECBCA is used to separate source signals from nent, η is the frequency index, and τ is the time. In order to extend SST,
audio signals mixed with various noises. The ECBCA is developed based a two-fold method was introduced [27]. A multivariate time-frequency
on WPT [26] and CBCA. An audio signal is first decomposed by WPT partitioning algorithm is used to partition the SST coefficients of the
into multiple wavelet sub-band signals. Then, CBCA is used to extract multi-channel data into multiple frequency bands. Then, a frequency
the sources from the wavelet sub-band signals. A mixed signal can be splitting process is conducted to separate the multivariate mono-com-
approximated by the following convolutive law: ponent signals. The MSST coefficients are determined by summing up
the multivariate instantaneous amplitude within each separated fre-
L−1
quency band.
x (t ) = ∑ A (l) s (t − l)
l=0 (1)
3.4. Adaptive kernel principal component analysis
Where, x(t ) = [x1, x2, …, x m]T are the m-channel sensor measurements,
s(t ) = [s1, s2, …, sn]T are n unknown sources, and After characterizing the audio signals with time-varying oscillatory
A (l) = [A1 (l), A2 (l), …, An (l)] is an m×n mixing-filter matrix in the lth components with MSST, multiple features, including eight statistical
frequency bin, and the order of the filter is L-1. Eq. (1) describes the features extracted in the time-frequency domain and an energy feature,
convolutive mixing processing in time domain. It can be rewritten in a are extracted. The AKPCA is used to convert the nine features into a set
∼ ∼
matrix form as x (t ) = A∼ sL , where A = [A (0), A (1), …, A ( L− 1)] and of linearly uncorrelated components. AKPCA is an extension of kernel
∼
sL = [s (k ), s (k − 1), …, s (k − L + 1)]. principal component analysis (KPCA). KPCA extracts principal compo-
Convolutive BSS is an effective tool to separate source signals from nents by mapping the input space into a feature space using a nonlinear
their mixture by identifying a separator matrix mapping function [29]. KPCA is not effective in processing non-sta-
B (l) = [B1 (l), B2 (l), …, Bn (l)] with an order of M-1. The separated tionary signals. Therefore, the AKPCA [30] was introduced by adap-
source signals y(t ) = [y1 , y2 , …, yn ]T are calculated by tively updating the kernel covariance matrix in KPCA.
For an original feature space f = (f1 , f2 , …, fn ) ∈ P p × n , with f1 ∈ P p
M−1
and p is the number of features. The PCA processing can be described
y (t ) = ∑ B (l) x (t − l)
l=0 (2) by Eq. (6).
∼∼ FF TW = λ W (6)
Eq. (2) can be rewritten as y (t ) = Bx M, where
∼ ∼ Where, F is the covariance matrix of f , and λ and W are the eigenvalues
B = [B (0), B (1), …, B (M − 1)] and xM = [x (k ), x (k − 1),
…, x (k − M + 1)]. Convolutive ICA has been used to determine the se- and eigenvectors of FF T , respectively. Then q PCs can be calculated
parator matrix. The CBCA has been used to relax the mutual-in- using the q-largest eigenvalues.
dependent assumption made in convolutive ICA by iteratively updating pc (q) = Wqf (7)
the separator matrix [23, 24]. The optimized separator matrix corre-
sponds to the maxima of the following objective function. Where, pc (q) is the q th PC, and Wq
are the eigenvectors corresponding
to the q-largest eigenvalues. In order to apply PCA to a nonlinear case,
∼ ( det(YyK ) )1/ K KPCA first projects f into a new space H via a mapping function ϕ ,
J (B) = then centers the projected data in H and subsequently conducts the
‖ρ (y (t ))‖nr (3)
traditional PCA processing on the centered data. The new covariance
Where, yK = [yK (K ), yK (K + 1), …, yK (K − M + 1)]T is a set of sepa- matrix in H is expressed as
rated sources, YyK is the covariance matrix of yK , ρ (y (t )) is the range of n
1
y(t ) , and ‖ρ (y (t ))‖nr is a measure of geometrics of the overall separated H= ∑ ϕ (fi ) ϕ (fi )T
sources (r ≥ 1), such as the volume of principal hyper-ellipse [24]. n i=1 (8)

68
Z. Li, et al. Journal of Manufacturing Processes 48 (2019) 66–76

Where, H is a kernel covariance matrix. The map ϕ adopts kernel to Table 1

calculate the inner products in H : k (fi , f j ) = ϕ (fi ) ∙ϕ (f j ) , where k (∙, ∙) Experimental design for milling tests.
is the kernel function (e.g., the radial basis function). Thus, Eq. (6) can Tool Condition Spindle Speed (rpm) Feed Rate (ipm)
be re-written as
Good 1400 14, 20, 30
Kn aq = nλq aq (9) 1800 9, 13, 17
2000 10, 15, 17.5
Where, Kn is the kernel gram matrix, λq and aq are the first q largest 2300 12, 18, 21
eigenvalues and the corresponding normalized eigenvectors of Kn , re- 2400 15, 25, 40
spectively. Then, for a new test data fn + 1, the PCs can be calculated as Slight 1400 14, 20, 30
1800 9, 13, 17
pcn + 1 (q) = aqϕ (fn + 1 ) . The kernel covariance matrix affects the effec- 2000 10, 15, 17.5
tiveness of PCs. AKPCA involves the following steps [30]: 2300 12, 18, 21
2400 15, 25, 40
(1) Construct and update the recursive formulation for KPCA by per- Average 1400 14, 20, 30
1800 9, 13, 17
forming the rotation and augmentation operations.
2000 10, 15, 17.5
(2) Estimate the kernel covariance matrix using a weighted sliding 2300 12, 18, 21
window. 2400 15, 25, 40
(3) Perform KPCA using the estimated kernel covariance matrix. Heavy 1400 14, 20, 30
1800 9, 13, 17
2000 10, 15, 17.5
Because the first two principal components usually contain up to 2300 12, 18, 21
85% of the original information [30], the first two principal compo- 2400 15, 25, 40
nents are selected, and then fed into machine learning algorithms. Severe 1400 14, 20, 30
1800 9, 13, 17
2000 10, 15, 17.5
4. Experimental setup 2300 12, 18, 21
2400 15, 25, 40
To demonstrate the proposed method, a set of milling tests with Failure 1400 14, 20, 30
1800 9, 13, 17
varying cutting conditions was conducted. All of the milling tests were 2000 10, 15, 17.5
conducted on a TRACK K3 EMX mill. The workpiece material is 6061 2300 12, 18, 21
aluminum. The flank wear of a set of end milling tools was character- 2400 15, 25, 40
ized using an optical microscope (Olympus BX60 M) according to ISO
8688-1 and ISO 8688-2. As shown in Fig. 2, tool wear conditions were
classified into six categories based on the average width of the flank

Fig. 2. Images of six tool wear classes: (a) Good: 0–20 μm in flank wear thickness, (b) Slight wear: 20–40 μm, (c) Average wear: 40–70 μm, (d) Heavy wear:
70–100 μm, (e) Severe wear: 100–150 μm, and (f) Failure: > 150 μm.

69
Z. Li, et al. Journal of Manufacturing Processes 48 (2019) 66–76

5. Results and discussions

5.1. Signal processing using ECBCA-MSST

The proposed ECBCA-MSST method was used to extract useful

features in the time-frequency domain from the multi-channel audio
signals. First, the audio signal collected from each microphone was
decomposed into four (4) wavelet sub-bands using WPT. Therefore,
twelve (12) sub-bands were generated from three microphones. The
BSS model was then used for denoising. Because the number of the
sources in the raw audio signals is unknown, an average signal to total
ratio (ASTR) is used to determine the number of source signals.
∑ El
ASTR =
Fig. 3. Experimental setup for the milling tests. EL (10)
Where, El is the energy of the lth (l = 1, 2, …, L. L is the number of
output sources) separated source and E is the energy of original three-
channel signals. A smaller ASTR indicates better denoising perfor-
mance. Fig. 4 shows the ASTR values of different output numbers in the
BSS separation process under the good condition (a spindle speed of
1800 rpm and a feed rate of 13 ipm). The smallest value of ASTR was
obtained when the number of BSS outputs is six. Similar results were
observed for other tool wear conditions. Therefore, the number of
source signals is six.
Fig. 5 compares the original time-frequency spectrum generated by
STFT against the time-frequency spectrum generated by ECBCA for the
good tool condition. As shown in Fig. 5(a), six principal components in
the time-frequency domain, including three components ([0 0.05],
[0.35 0.4], and [0.75 0.76]) in the frequency domain and three com-
ponents ([7.5 s 10 s], [11 s 12 s], and [12.5 s 14 s]) in the time domain,
were decomposed from the raw signal. The three components in the
Fig. 4. The values of ASTR under the good tool wear condition at a rotation frequency domain are due to three sources of vibration generated
speed of 1800 rpm and a feed rate of 13 ipm. during the milling process. The three components in the time domain
are due to three sources of sound waves generated during the milling
process. Fig. 5(b) shows strong background noise in red. ECBCA re-
wear land of two flutes. For each flute, the width of flank wear was
moved some of the noise while retaining six principal components.
determined by taking an average of ten measurements at different lo-
In order to extract useful information about tool wear in the time-
cations of the cutting edge.
frequency domain, the ECBCA output was used as the input of the MSST
Table 1 lists the details about the cutting conditions. Each tool wear
processor. Figs. 6–11 show the time-frequency characteristics of the six
category involves five different cutting speeds and three different feed
different tool wear conditions by using MSST. The SST analysis results
rates. To collect audio signals during milling tests under varying cutting
were also presented for comparison purpose. The spindle speed was
conditions, three condenser microphones (ECOOPRO EO-200) were
1800 rpm (i.e., spindle angular frequency fr =30 Hz) for each condi-
used to collect real-time audio signals from different directions and
tion. Theoretically, the milling dynamics model is analogical to the
distances as shown in Fig. 3. The sampling frequency of the sound
dynamics of a rotor rubbing model, where fr and its harmonics such as 5
collection was set to 44.1 kHz. More details about the experiments can
or 10 times harmonics (i.e. 150 Hz or 300 Hz) would be the dominating
be found in [16].
frequency components in the time-frequency domain [31]. Similarly, in
the time-frequency plot of the tool cutting dynamics, dominant energy
strips would appear at the harmonics of the spindle rotational

Fig. 5. Time-frequency spectrum generated by (a) STFT and (b) ECBCA.

70
Z. Li, et al. Journal of Manufacturing Processes 48 (2019) 66–76

Fig. 6. Time-frequency representation under the good condition: (a) MSST and (b) SST.

Fig. 7. Time-frequency representation under the slightly tool wear condition: (a) MSST and (b) SST.

Fig. 8. Time-frequency representation under the average tool wear condition: (a) MSST and (b) SST.

frequency fr. prominent in the time-frequency spectrum of the tool-workpiece dy-

During the milling tests, the spindle rotational frequency (fr namics under six different tool health conditions. It should be noted
=30 Hz) and the corresponding harmonic frequencies are dominant in that the amplitude of the cutting dynamics varies with different tool
the lower frequency band of the audio sound, which is consistent with health conditions and increases with the severity of the tool wear. The
the observations in previous researches [31–33]. Because the main amplitude of good tool is the smallest while the severe tool wear and
exciting force in the milling operation is the contact between the cut- tool failure generate the largest amplitude. In addition, when com-
ting tool and workpiece, the dynamical responses of the milling ma- paring the Figs. 6(a)–11 (a) with Figs. 6(b)–11 (b), we observed that the
chine with good and worn tools are similar. The health condition of the background noise has been significantly removed by MSST. As a result,
tool affects the energy distribution of the tool-workpiece dynamics. As the signal processing results in Figs. 6–11 demonstrate the effectiveness
shown in Figs. 6(a)–11 (a), the energy distribution at about 5, 10, 35 of the proposed method. It should be noted that the energy magnitude
and 40 times harmonics of the spindle rotational frequency are of different tool wear states in Figs. 6(a)–11 (a) are very different.

71
Z. Li, et al. Journal of Manufacturing Processes 48 (2019) 66–76

Fig. 9. Time-frequency representation under the heavy tool wear condition: (a) MSST and (b) SST.

Fig. 10. Time-frequency representation under the severe tool wear condition: (a) MSST and (b) SST.

Fig. 11. Time-frequency representation under the failure condition: (a) MSST and (b) SST.

Therefore, the sum of energy amplitudes in the low frequency band 5.2. Feature extraction
(0–1400 Hz) was used as a feature for tool wear classification. Fig. 12
shows the zoomed figures of Figs. 6(a)–11 (a) in the range of The audio signal was collected within a time interval of about 30 s
0–1400 Hz. The average values of energy sum features calculated from during each cutting test. In order to generate sufficient training and
15 samples of each tool wear condition are 0.4967, 2.2800, 8.0493, testing datasets, the recorded sound wave in each cutting test was
1.3047, 3.0647, and 17.3893. The average value of energy sum in good segmented into one-second sub-datasets. After the segmentation, there
tool condition is the smallest while the failure condition produces the were 333, 351, 407, 439, 627, and 500 sub-datasets under good, wear,
largest average value. average wear, heavy wear, severe wear, and failure conditions, re-
spectively. 2657 sub-datasets were generated.
Eight (8) features (i.e., max, min, mean, median, moment, skewness,
kurtosis, and standard deviation) were extracted in the time-frequency

72
Z. Li, et al. Journal of Manufacturing Processes 48 (2019) 66–76

Fig. 12. Time-frequency spectrums under (a) good tool, (b) slight wear, (c) average wear, (d) heavy wear, (e) severe wear, (f) failure conditions.

domain for each sub-dataset using Eq. (11). nine (9) features were calculated for each sub-dataset and the total is
F9×2657. AKPCA was then employed to fuse the original nine features
⎧ fi = gi (e) into two new features. Thus, a new feature space H2×2657 was obtained.
8700 ∼k
⎨ ej = ∑ Ej Fig. 13 shows the AKPCA feature distribution of half of H. The first
⎩ k=0 (11) AKPCA feature (red dots) can be used for classification of most sub-
Where, fi (i = 1, 2, …, 8) denotes one of the eight (8) features and gi(∙) is datasets even if there are overlaps between slight and average wear and
the corresponding calculation function (i.e. the functions of max, min, between good and heavy wear conditions. The second AKPCA feature
mean, median, moment, skewness, kurtosis, and standard deviation), e (blue circles), however, can provide additional information to distin-
= [e1, e2, …, ej] (j = 1, 2, …, N and N is the number of sample points of guish the slight and average wear conditions. The fused features were
each sub-dataset) is the time series of the energy-amplitude sum at each fed into machine learning algorithms in order to classify tool wear
∼k conditions.
sample point along the frequency axis, E j denotes the energy-ampli-
th
tude at j sample point and k (∈ R ) Hz frequency and 8700 (Hz) is the
maximum frequency value of each audio signal in the milling tests. 5.3. Classification of tool wear conditions
In addition, an energy feature was extracted by summing up the
energy-amplitude in the range of 0–1400 Hz as shown in Fig. 12. Thus, Four classifiers, including CART, random forest (RF), kNN and SVM,

73
Z. Li, et al. Journal of Manufacturing Processes 48 (2019) 66–76

Fig. 13. The first two principal components decomposed by AKPCA.

Table 2
Evaluation of the proposed method on 50% dataset.
Condition Classification Accuracy (%)

CART RF kNN SVM

Without ECBCA With ECBCA Without ECBCA With ECBCA Without ECBCA With ECBCA Without ECBCA With ECBCA

Good 96.39 98.19 84.34 86.14 94.58 95.78 98.19 100

Slight 92.57 94.29 85.71 88.00 96.00 97.14 97.14 98.29
Average 96.55 98.52 97.54 98.03 97.54 98.52 96.06 98.03
Heavy 95.43 98.17 87.67 89.50 98.63 99.54 96.80 98.63
Severe 98.73 99.36 99.04 99.36 98.09 98.66 96.24 98.66
Failure 92.00 93.20 98.40 99.36 94.40 97.20 96.00 97.20
Overall Accuracy 95.28 96.96 92.12 93.40 96.54 97.81 96.74 98.47

Table 3 Table 5
Comparisons of SVM monitoring results with different signal processing Monitoring performance of RF using different training size.
methods.
Size of the training dataset
Results in [16] SST MSST ECBCA-MSST
50% 60% 70% 80% 90%
Good 99.4% 84.34% 98.19% 100%
Slight 97.5% 77.71% 97.14% 98.29% Good 86.14% 82.17% 86.00% 94.03% 90.91%
Average 90.7% 61.08% 96.06% 98.03% Slight 88.00% 86.43% 89.52% 91.43% 94.29%
Heavy 92.4% 66.21% 96.80% 98.63% Average 98.03% 99.39% 98.36% 100% 100%
Severe 98.5% 88.54% 96.24% 98.66% Heavy 89.50% 94.89% 97.73% 93.18% 97.73%
Failure 97.0% 81.20% 96.00% 97.20% Severe 99.36% 99.15% 100% 100% 100%
Overall Accuracy 95.92% 76.51% 96.74% 98.47% Failure 99.36% 92.50% 92.67% 86.00% 96.00%
Overall Accuracy 93.40% 92.51% 94.05% 94.11% 96.49%

Table 4
Monitoring performance of CART using different training size. Table 6
Monitoring performance of kNN using different training size.
Size of the training dataset
Size of the training dataset
50% 60% 70% 80% 90%
50% 60% 70% 80% 90%
Good 98.19% 96.99% 97.00% 100% 93.94%
Slight 94.29% 95.00% 97.14% 94.29% 100% Good 95.78% 96.99% 98.00% 98.51% 100%
Average 98.52% 100% 99.18% 100% 100% Slight 97.14% 95.00% 99.05% 98.57% 97.14%
Heavy 98.17% 98.86% 96.21% 94.32% 100% Average 98.52% 98.16% 100% 99.18% 98.50%
Severe 99.36% 99.15% 100% 100% 100% Heavy 99.54% 99.18% 99.24% 98.86% 99.33%
Failure 93.20% 96.50% 94.67% 99.00% 100% Severe 98.66% 100% 98.86% 98.15% 99.00%
Overall Accuracy 96.96% 97.75% 97.37% 97.93% 98.99% Failure 97.20% 97.56% 100% 100% 96.00%
Overall Accuracy 97.81% 97.97% 98.94% 98.85% 98.45%

were used to identify tool wear conditions. Table 2 lists the tool wear
monitoring results using half of the feature space H as the training randomly selected for the classifiers and the training-test procedure
dataset and the other half as test dataset. The training units were repeated five times. In the training process of the classifiers, 500 trees

74
Z. Li, et al. Journal of Manufacturing Processes 48 (2019) 66–76

Table 7
Monitoring performance of SVM using different training size.
Size of the training dataset

50% 60% 70% 80% 90%

Good 100% 97.74% 99.00% 97.01% 100%

Slight 98.29% 97.14% 95.24% 97.14% 100%
Average 98.03% 98.77% 100% 100% 100%
Heavy 98.63% 100% 97.73% 100% 100%
Severe 98.66% 99.18% 100% 98.15% 100%
Failure 97.20% 96.50% 96.67% 99.00% 94.00%
Overall Accuracy 98.47% 98.22% 98.11% 98.55% 99.00%

were used for RF, k was chosen as 15 for kNN, and the radial basis
function with a width parameter 0.5 was used for SVM. The average
classification accuracy for five tests is listed in Table 2. For the ECBCA
denoising performance, the classification accuracy for each tool wear Fig. 15. Overall classification accuracy of four classifiers with varying training
condition was improved. In addition, the overall classification accuracy size.
for the six tool conditions using ECBCA-based methods is higher than
that without ECBCA, and the best result is obtained by the SVM clas- the BSS-based multi-channel signal pre-processing. Furthermore, the
sifier, 98.47% overall. MSST performed better than the frequency-domain method in [16] with
In a previous study [16], SVM classifier was employed to identify respect to recognition accuracy. This is because MSST is able to extract
the six tool wear conditions with an overall accuracy of 95.92%. In this useful features by analyzing multiple signals in the time-frequency
study, the frequency domain analysis is extended into the time-fre- domain. It should be also noted that the SST method is unsuitable for
quency domain. The tool wear monitoring performance using an SVM the tool wear monitoring in this study due to ineffective recognition
classifier with the signal processing method in [16] and the proposed results.
ECBCA-MSST method is compared in Table 3. It can be seen that the
proposed method achieved a 100% accuracy in identifying the good
tool condition against the other three methods. The monitoring accu- 5.4. Discussions
racy of the proposed method for the other five tool wear conditions was
also higher. The main reason for the improvement is the contribution of We have demonstrated the proposed method for tool wear mon-
itoring and prediction using half of H as the training dataset. To

Fig. 14. Tool wear monitoring performance using different training size: (a) CART, (b) RF, (c) kNN & (d) SVM.

75
Z. Li, et al. Journal of Manufacturing Processes 48 (2019) 66–76

evaluate the effect of training size on the performance of the predictive [8] Wu D, Jennings C, Terpenny J, Gao RX, Kumara S. A comparative study on machine
model, a varying amount of H was used for training and validation. learning algorithms for smart manufacturing: tool wear prediction using random
forests. J Manuf Sci Eng 2017;139(7):71018.
Tables 4–7 and Fig. 14 show the prediction accuracy of the predictive [9] Weller EJ, Schrier HM, Weichbro B. What sound can Be expected from a worn tool. J
models trained by CART, RF, kNN, and SVM. As expected, as the Inst Eng 1969;91(3). 525-&.
amount of training data increases, the classification accuracy increases. [10] Mannan M, Kassim AA, Jing M. Application of image and sound analysis techniques
to monitor the condition of cutting tools. Pattern Recognit Lett
Fig. 15 shows the overall classification accuracy of four classifiers 2000;21(11):969–79.
with differing training size. As the training data size increases, the [11] Delio T, Tlusty J, Smith S. Use of audio signals for chatter detection and control. J
overall classification accuracy increases. The best overall accuracy, Eng Ind-T Asme 1992;114(2):146–57.
[12] Salgado D, Alonso F. An approach based on current and sound signals for in-process
99.0%, was obtained by SVM utilizing 90% of the total dataset. tool wear monitoring. Int J Mach Tools Manuf 2007;47(14):2140–52.
[13] Aliustaoglu C, Ertunc HM, Ocak H. Tool wear condition monitoring using a sensor
6. Conclusions and future work fusion model based on fuzzy inference system. Mech Syst Signal Process
2009;23(2):539–46.
[14] Ubhayaratne I, Pereira MP, Xiang Y, Rolfe BF. Audio signal analysis for tool wear
A novel audio-based signal processing approach was presented to monitoring in sheet metal stamping. Mech Syst Signal Pr 2017;85:809–26.
improve the accuracy of tool wear prediction. The extended con- [15] Seemuang N, McLeay T, Slatter T. Using spindle noise to monitor tool wear in a
volutive bounded component analysis (ECBCA) was used to process turning process. Int J Adv Manuf Tech 2016;86(9–12):2781–90.
[16] Kothuru A, Nooka SP, Liu R. Application of audible sound signals for tool wear
multivariate audio signals in order to separate a set of source signals monitoring using machine learning techniques in end milling. Int J Adv Manuf
from a set of mixed signals. The multivariate synchrosqueezing trans- Technol 2018;95(9–12):3797–808.
form (MSST) was used to characterize multivariate audio signals with [17] Lauro C, Brandão L, Baldo D, Reis R, Davim J. Monitoring and processing signal
applied in machining processes–A review. Measurement 2014;58:73–86.
time-varying oscillatory properties. The adaptive kernel principal [18] Zhu K, San Wong Y, Hong GS. Wavelet analysis of sensor signals for tool condition
component analysis (AKPCA) was used to transform the denoised sig- monitoring: a review and some new results. Int J Mach Tools Manuf
nals into a feature space. A set of cutting tests was performed to collect 2009;49(7–8):537–53.
[19] Fang N, Pai PS, Mosquea S. Effect of tool edge wear on the cutting forces and vi-
audio signals using multi-channel microphones. These audio signals brations in high-speed finish machining of Inconel 718: an experimental study and
were used for training and validating the predictive model. The ex- wavelet transform analysis. Int J Adv Manuf Technol 2011;52(1–4):65–77.
perimental results have shown that the proposed ECBCA-MSST method [20] Cao H, Lei Y, He Z. Chatter identification in end milling process using wavelet
packets and Hilbert–Huang transform. Int J Mach Tools Manuf 2013;69:11–9.
was able to extract key audio sound sources that are closely related to [21] Zhu K, Hong GS, Wong YS, Wang W. Cutting force denoising in micro-milling tool
the dynamic response of tool milling operations. In addition, the AKPCA condition monitoring. Int J Prod Res 2008;46(16):4391–408.
method was able to improve prediction accuracy while reducing the [22] Shi XH, Wang R, Chen QT, Shao H. Cutting sound signal processing for tool
breakage detection in face milling based on empirical mode decomposition and
dimensionality of the extracted features. In the future, the proposed
independent component analysis. J Vib Control 2015;21(16):3348–58.
approach will be applied to large volumes of real-time audio signals. [23] Inan HA, Erdogan AT. A convolutive bounded component analysis framework for
potentially nonstationary independent and/or dependent sources. IEEE T Signal
Acknowledgements Proces 2015;63(1):18–30.
[24] Inan HA, Erdogan AT. Convolutive bounded component analysis algorithms for
independent and dependent source separation. IEEE Trans Neural Netw Learn Syst
The research reported in this paper is partially supported by the 2015;26(4):697–708.
University of Central Florida (UCF). Any opinions, findings, and con- [25] Li Z, Yan X, Wang X, Peng Z. Detection of gear cracks in a complex gearbox of wind
turbines using supervised bounded component analysis of vibration signals col-
clusions or recommendations expressed in this paper are those of the lected from multi-channel sensors. J Sound Vib 2016;371:406–33.
authors and do not necessarily reflect the views of the UCF. [26] Tazebay MV, Akansu AN. Adaptive subband transforms in time-frequency excisers
for dsss communications-systems. IEEE T Signal Proces 1995;43(11):2776–82.
[27] Ahrabian A, Mandic DP. A class of multivariate denoising algorithms based on
References synchrosqueezing. IEEE T Signal Proces 2015;63(9):2196–208.
[28] Meignen S, Oberlin T, McLaughlin S. A new algorithm for multicomponent signals
[1] Wu D, Jennings C, Terpenny J, Kumara S, Gao RX. Cloud-based parallel machine analysis based on SynchroSqueezing: with an application to signal sampling and
learning for tool wear prediction. J Manuf Sci Eng 2018;140(4):41005. denoising. IEEE T Signal Proces 2012;60(11):5787–98.
[2] Ji W, Shi J, Liu X, Wang L, Liang SY. A novel approach of tool wear evaluation. J [29] Pokharel PP, Liu WF, Principe JC. Kernel least mean square algorithm with con-
Manuf Sci Eng 2017;139(9):91015. strained growth. Signal Process 2009;89(3):257–65.
[3] Gao R, Wang L, Teti R, Dornfeld D, Kumara S, Mori M, et al. Cloud-enabled prog- [30] Ding MT, Tian Z, Xu HX. Adaptive kernel principal component analysis. Signal
nosis for manufacturing. CIRP Ann Manuf Technol 2015;64(2):749–72. Process 2010;90(5):1542–53.
[4] Wang J, Wang P, Gao RX. Enhanced particle filter for tool wear prediction. Int J Ind [31] Hu A, Xiang L, Zhang Y. Experimental study on the intrawave frequency modulation
Manuf Syst Eng 2015;36:35–45. characteristic of rotor rub and crack fault. Mech Syst Signal Process
[5] Snr DED. Sensor signals for tool-wear monitoring in metal cutting operations—a 2019;118:209–25.
review of methods. Int J Mach Tools Manuf 2000;40(8):1073–98. [32] Reju VG, Koh SN, Soon Y. Underdetermined convolutive blind source separation via
[6] Zhang J, Starly B, Cai Y, Cohen PH, Lee Y-S. Particle learning in online tool wear time–frequency masking. IEEE Trans Audio Speech Lang Process
diagnosis and prognosis. J Manuf Process 2017;28:457–63. 2010;18(1):101–16.
[7] Sick B. On-line and indirect tool wear monitoring in turning with artificial neural [33] Peng D, Smith WA, Randall RB, Peng Z. Use of mesh phasing to locate faulty planet
networks: a review of more than a decade of research. Mech Syst Signal Process gears. Mech Syst Signal Process 2019;116:12–24.
2002;16(4):487–546.

76