Tool Wear Condition Monitoring Using Emitted Sound Signals By Simple
Machine Learning Technique
1C.LOGESH PERUMAL, 1S.B.BHADRINATHAN, 2ANDREWS SAMRAJ
1Department of Artificial Intelligence and Data Science, Mahendra Engineering College,
Namakkal,Tamil Nadu, INDIA
2Department of Computer Science and Engineering, Mahendra Engineering College
Namakkal,Tamil Nadu, INDIA
Abstract— as a continuous enhancement to the tool wear monitoring using non-disturbing method of
sound wave analysis, a simple machine learning technique enhances the prediction to better levels and
reduces the procedures. A simple linear regression Algorithm was used to train and predict the trends of
various degrees of tool wear to distinguish them from each other. The results based on this simple linear
regression were successful in showing the difference of sound patterns and are reported.
Keywords— Tool wear monitoring, industry 4.0, non-disturbing testing, Machine Learning.
Received: July 15, 2021. Revised: April 8, 2022. Accepted: May 10, 2022. Published: June 1, 2022.
1. Introduction
Using the pressure difference in sound that is arising from an
ongoing machining job to find the degree of tool wear is a
good method followed by the recent advancements in
automation. As the Industry 4.0 compliance giving
continuous demands for machine learning and perfect
automation, there is always an opportunity for improvement
in this work. This challenging task of knowing the degree of
tool wear without stopping the process was dealt differently
by different researchers. Usually the conventional methods
fix sensors and cameras to get tool wear conditions and the
parameters drawn are used to decide the action. This
happens in the production process using CNC machine that
highly depends on the tool condition involved in the job
work. It is important to ensure the good quality of tool to get
the genuine quality in end product. In such turning process
an automated conditional monitoring of an active and
effective tool, involved in drilling and milling requires a
mechanism that works without hindrance of the operation
sequence is highly productive.
A very useful method of drawing sound information
from the evaluated feed motor current, force of feed, cutting,
are the parameters suggested by Alonso et al. [3] suggested
a guessing method of tool flank wear a ANN. The
directional proportional connection between the maximum
extend of vibration with the increasing tool wear was found
by Sadettin et al [4]. The sound that is created during the
machine cutting process is used as the unique parameter for
monitoring the degree of tool flank wear from the research of
Ming-Chyuan et al [5] and Alonso F.J et al. [6]. A much
simpler and cost effective mechanism following the methods
of [5] and [6] is developed by sound converting process to
estimate the tool flank wear swiftly via dynamic assessment
group technique through just a portion of discharged noise
instead of a big piece of sound wave by A Samraj et al. [7].
The distinctive signal is handled in the form of linear and not
permanently immobile; or else the result deriving Fourier
spectrum may produce significantly lower physical sense
was suggested by Peng et al [8]. In following the Fourier
transformation leads the world wide properties of the signal
rather than local properties according to the direction of
Huang et al [9]. Hilbert-Huang transformation also helps to
estimate the results to a good extent.[10] A continuous
enhancement of the process leads to much simpler method
and effective results and is presented by Prakash & A.Samraj
[11,12,13]
Figure1: Different types of tool bits used in the cutting
process
In alignment to Industry 4.0 automation a Machine learning
approach seems to be appropriate to address this problem.
Hence we started a basic ML approach using a simple Linear
Regression and found the approach results in fruitful but
different way of representations.
This initiative is to change the direction of tool
wear monitoring from conventional and basic techniques to
advanced and compatible technique of ML. A refined
approach which will be more adoptive to any Industrial 4.0
systems would be the ML technique. Here we started it with
the first ML technique on the emitted sounds and found it as
a competent method that augments any efficient automated
manufacturing. The old alternative such as image capturing
in videos and other monitoring devices causes expenses and
erroneous predictions. In this paper we presented the
possible results with the estimation of wear in single point
cutting tools by the displays of Linear Regression
estimations and their convergence styles.
DESIGN, CONSTRUCTION, MAINTENANCE
DOI: 10.37394/232022.2022.2.22
C. Logesh Perumal, S. B. Bhadrinathan, Andrews Samraj
E-ISSN: 2732-9984
168
Volume 2, 2022
2. Methodoloy
2.1 Eqipment and setup
The setting up of equipment for this sound capturing
experiment due to the noise created out of the vibration
involve a sensitive micro phone which has a size of ¼” in
diameter and its market identification name is PCB 130 D20.
The capacity of this microphone is to record a vital extent of
noise till 12dB. Hence PCB 130 D20 becomes the most
appropriate microphone to capture and record the sound
created during turning process with a reaction to the
frequency between the extend of 20Hz to 20 kHz with the
precision of noise variations of ±0.5dB. This microphone is
using a BNC connector and a high temperature resistant
material made by polymer, for removing the requirement of
external polarization, which include freeze concerned with
charges, implemented on the back platys top. In general this
PCB 130 D 20 microphone has records of being widely used
in the measuring of noise in multi-channel machinery as well
as sound power measurements.
The microphone and its arrangement with position and
connection for measuring sound intensity are shown in
Figure. 2 suitable to be deployed in the machining process.
.
Figure. 2 The Microphone used to record the cutting
sound and its setup
2.2 Simple Linear Regression:
One of the basic ML technique adopted is this statistical
model to predict the relationship between independent and
dependent variable. Linear regression has to be applied on a
1000 point statistical model of the acoustic signal captured
during the machining process with Fresh, Slightly worn tool
and Severely worn tool as three different categories.
The simple linear regression is to find the relationship
between two variables with a linear trend of the future
values of these variables. So the regression is a process of
estimating the value of one variable using another variable.
2.3 The mechanics
A substantial quantity of sound is constantly produced
during the turning process due to the vibration produced
from the work machine tool and the work-piece. This
produced noise during the machining process is anticipated
according to the intensity and the size of surface of contact
to which the insertion of cutting flank were occurred. Along
with that the other assorted vibrations produced from the
tool helps us in the tool wear estimation process. But some
vibration interruptions from the environment spoil the
quality. Hence they should be separated using appropriate
filters since they appear in the less than average frequency
ranges of null point [0] and 2 kHz. But the effect of the
conversion procedure is important when these interruptions
over the 2 kHz frequency measure.
Out of many materials like Steel Aluminum, and alloy,
the Aluminum work-piece material is selected for this
experiment. The work piece was having a diameter around
50 mm. The cutting insert tool bit used here was a carbide
insert NR9. Throughout this experiment the micro phone’s
output varies according to the sounds pressure on the time
discipline proportionally changes.
A PCB 130 D20 associated software was used for the
recordings, which has the commercial name called ‘Gold
Wave’ , and was very helpful to do the recoding of sound
signal with the modal rate of 44100 points/sec using the pre
polarized condenser microphone. The sampling rate of
44100/sec is high and the resulted data due to this high
sampling rate is a challenging and is an over head for
processing. We designed the Experiments with few constant
cutting parameters like depth of the cut, cutting speed as well
as feed rate and kept them the same throughout the
experiment. We recorded emitted sound waves during 15
trials during the turning operation. Three sessions of
recordings were done for all the three status conditions of the
aluminum tool bit. (Fresh tool without any flank wear- 15
trials, slightly worn tool – 15 trials which is having an
approximately 0.2- 0.25 millimeter flank wear also severely
worn out tool of approximately 0.4 millimeter flank wear –
15 trials). A free run sound was also recorded to have it as a
base reference in the start of the experiment while the
machine is operated freely without touching the work piece
by the tool bit. This particular measurement is then labeled as
no control run. Subsequently tools of experiments like Fresh
tool, slightly worn tool and severely worn tool were used in
recordings and are labeled it their names after the sound
waves are recorded.
2.4 Experiments with 1000 data points of the
sound signal
The data involved in this experiment contains a sequence of
1000 data points of acoustic signal for 3 different categories.
The signal recorded during the job work when a fresh tool bit
was used, when a slightly worn tool bit was used and a
severely worn tool bit was used are the three different data
used in this experiment.
The sequence of 1000 data values are converted from 1000 X
1 array form to 10 X 100 array form by splitting the data in
DESIGN, CONSTRUCTION, MAINTENANCE
DOI: 10.37394/232022.2022.2.22
C. Logesh Perumal, S. B. Bhadrinathan, Andrews Samraj
E-ISSN: 2732-9984
169
Volume 2, 2022
to 100 data points. Then the dataset posses the dependent and
independent variables ‘x’ and ‘y’ respectively with the time
stamps. Then once again these 10 X 100 data set array is split
into training and testing set , by using the ML framework
software as the training size of 70% of the given data and
30% as the test data out of the 1000 data points.
The next step is to build the ML model by using Linear
Regression function. Create an object of ML as ‘regress’ to
fit the trained data as x_train and y_train in the regress
object.
Where ‘regress’ object is representing the linear regression
of the training and test data using the function. The
prediction based on the training occurs as the best fitting line
on the given test data. To visualize the output, the predictions
are plotted with variables x_train and y_train as the x and y
axis of the graph along with the predicted best fitting
regression line of test data.
3. Results and Findings
Fresh Data:
The plot in figure 3 shows the time in the x axis and
wave pressure value in the y axis plotted into the graph. It
consists 1000 Data points out of which a random 70% (700)
is chosen as a training set, and the remaining 30% (300) is
the test set. The regression line is plotted for the test set and
the predicted values show the regression line.
Figure 3: Linear Regression of 1000 data points in sound
signal from Fresh category Tool.
The plot in figure 4 shows the time in the x axis and
wave pressure value in the y axis plotted into the graph. But
it consists 300 Data points out of which a random 70% (210)
is chosen as a training set, and the remaining 30% (90) is the
test set. The regression line is plotted for the test set and the
predicted value shows the regression line.
Figure 4 : Linear Regression of 300 data points in
sound signal from Fresh Category Tool.
Severe Data:
The plot in figure 5 with same axis legends is showing
the change in wave pressure value depicted in the y axis of
the graph. It is due to the worn out tool that is being used
changes the aquostic noise. Here too a 1000 data points out
of which a random 70% (700) is chosen as a training set, and
the remaining 30% (300) is the test set. The regression line is
plotted for the test set (30%) and the predicted values show
the regression line.
Figure 5: Linear Regression of 1000 data points in sound
signal from Severe Category Tool.
The plot in figure 6 shows the time in x axis and wave
pressure value in the y axis plotted into the graph. It contains
300 Data points out of which a random 70% (210) is chosen
as a training set, and the remaining 30% (90) is the test set.
The regression line is plotted for the test set and the predicted
values show the regression line..
DESIGN, CONSTRUCTION, MAINTENANCE
DOI: 10.37394/232022.2022.2.22
C. Logesh Perumal, S. B. Bhadrinathan, Andrews Samraj
E-ISSN: 2732-9984
170
Volume 2, 2022
Figure 6 : Linear Regression of 300 data points in sound
signal from Severe Category Tool.
Slight Data :
The plot in figure 7 shows the time in the x axis and
wave pressure value in the y axis plotted into the graph. It
consists 1000 Data points out of which a random 70% (700)
is chosen as a training set, and the remaining 30% (300) is
the test set. The regression line is plotted for the test set and
the predicted values show the regression line.
Figure 7 : Linear Regression of 1000 data points in sound
signal from Slight Category Tool.
The plot in figure 8 shows the time in the x axis and
wave pressure value in the y axis plotted into the graph. But
it consists 300 Data points out of which a random 70% (210)
is chosen as a training set, and the remaining 30% (90) is the
test set. The regression line is plotted for the test set and the
predicted values shows the regression line.
Figure 8 : Linear Regression of 300 data points in sound
signal from Slight Category Tool.
4. Results & Discussions
The identification of degree of tool wear can be identified
by the proposed machine learning technique without any
complex transformations. The signal samples taken over the
period of time is reduced and analysed for the quality of
prediction and found unaffected. It means that a very small
piece of sound signal is enough to identify the degree of tool
wear by this technique. There are some specific features
patterns are found for different category of the signals which
are clearly grouped down by our simple ML techniques. The
accumulation of processed data points over a particular area
can be found in the figures 3 to 8 . Especially in figures 5
and 6 the intensification of data points near 0 could be
found.
5. Conclusion
The simple features constructed using simple machine
learning technique used for identifying the degree
of tool wear is found suitable in different
categories. On observing the table 1 the change in
linear regression for fresh , slight and severe tool
categories are found in perfect intonations. It is
found that the proposed approach is simple, easier
and does not involve any complex mathematical
derivations or transformations. The features based
on the sound pressure represented by the data
points are only considered for decision making.
This, valid contribution to this particular research
area will definitely contribute to save cost and
time.
DESIGN, CONSTRUCTION, MAINTENANCE
DOI: 10.37394/232022.2022.2.22
C. Logesh Perumal, S. B. Bhadrinathan, Andrews Samraj
E-ISSN: 2732-9984
171
Volume 2, 2022
Table 1: Start and end points of linear regression for three Different categories
References
[1] E. Dimla, and S. Dimla , “Sensor signals for tool-wear
monitoring in metal cutting operations-a review of methods”,
International Journal of Machine Tools & Manufacture, vol.
40, pp. 1073- 1098, 2000.
[2] R. G. Silva, R. L. Reuben, K. J. Baker, and S. J. Wilcox,
“Tool wear monitoring of turning operations by neural
network and expert systemClassification of a feature set
generated from multiple sensors,” Mechanical Systems and
Signal Processing, vol. 12, pp. 319–332, November 1998.
[3] F. J. Alonso and D. R. Salgado, “Application of singular
spectrum analysis to tool wear detection using sound signals,”
Proc. IME, Part B: Journal of Engineering Manufacture, vol.
219, no. 9, pp. 703-710, 2005
[4] Sadettin, O., S. Orhan, A. Osman Er, N. Camus-cu, and E.
Aslan, “Tool wear evaluation by vibration analysis during end
milling of AISI D3 cold work tool steel with 35 HRC
hardness,” NDT&E International, vol. 40, pp. 121–126, 2007.
[5] M. Chyuan Lu, E. Kannatey-Asibu, “Analysis of Sound
Signal Generation Due to Flank Wear in Turning,” Journal of
Manufacturing Science and Engineering, vol. 124, no. 4,
doi:10.1115/1.1511177, pp. 799-808 , November 2002.
[6] F. J. Alonso and D. R. Salgado, “Application of singular
spectrum analysis to tool wear detection using sound signals,”
Proc. IME, Part B: Journal of Engineering Manufacture, vol.
219, no. 9, pp. 703-710, 2005.
[7] A. Samraj, S. Sayeed, J. Emerson Raja, J. Hossen, and A.
Rahman, “Dynamic Clustering Estimation of Tool Flank
Wear in Turning Process using SVD Models of the Emitted
Sound Signals,” World Academy of Science, Engineering and
Technology, vol. 80, pp.
1322-1326, 2011.
[8] Z. K. Peng, Peter W. Tse, and F. L. Chu, “An improved
Hilbert-Hung transform and its application in vibration signal
analysis,” Journal of Sound and Vibration, vol. 286, pp. 187-
205, August 2005.
[9] N. E. Huang, Z. Shen, S. R. Long, M. L. C. Wu, H. H. Shih,
Q. N. Zheng, N. C. Yen, C. C. Tung and H.H. Liu, “The
empirical mode decomposition and the Hilbert spectrum for
nonlinear and non-stationary time series analysis,” in Proc.
Royal Society of LondonSeries A - Mathematical Physical
and Engineering Sciences, 1998, pp. 903-995.
[10] A. G. Rehorn, J. Jiang, and P. E. Orban, "State-of the art
methods and results in tool conditioning monitoring: a
review", International Journal of Advanced Manufacturing
Technology,vol.26,pp.693-710,2005.
[11] Bainian Li, Kongsheng Zhang, and Jian Xu, "Similarity
measures and weighed fuzzy c-mean clustering algorithm",
International Journal of Electrical and Computer Engineering, vol.
6, no. 1, pp. 1-4. 2010.
[12] K.Prakash, Andrews Samraj, “ Tool Flank Wears Estimation
by Simplified SVD on Emitted Sound Signals”, IEEE Conference
on Emerging Devices and Smart Systems, Proceedings, India,
2017.
[13] K.Prakash, Andrews Samraj, “Tool Wear Condition
Monitoring using Acoustic Analysis of Emitted Sound Signals by
Peak to Peak Analysis”, ISERD 93rd International Conference,
Hanoi, Vietnam 2017.
data
fresh
slight
start
end
start
end
start
end
1000
0 , -1.25
1000 , 0.4
0 , -1.25
1000 , 0.2
0 , -0.2
1000 , 0.0
300
700 , -0.12
1000 , 0.3
700 , -0.22
1000 , 0.25
700 , 0
1000 , 0.0
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the Creative
Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en_US
DESIGN, CONSTRUCTION, MAINTENANCE
DOI: 10.37394/232022.2022.2.22
C. Logesh Perumal, S. B. Bhadrinathan, Andrews Samraj
E-ISSN: 2732-9984
172
Volume 2, 2022