Single point tool wear condition monitoring with acoustic signal and radial

basis function neural network

M.Gopi Krishnaa* C.Ramesh kannanb J Emerson Rajac, and
S Purushothamand
M.E student
Engineering design
Department of mechanical Engineering
University college of Engineering,
Anna university, Madurai
[email protected]

Faculty, Department of mechanical Engineering

University college of Engineering,
Anna university, Madurai
Email id: [email protected]
c
Faculty of Engineering and Technology
Multimedia University, Malaysia
Tel : 606 2523235. Fax 606 2316552
Email : [email protected]

D
Professor, PET Engineering College, India-627117,
[email protected]
d
Email :[email protected]

Abstract-This paper presents single point Tool wear monitoring method. It has been implemented by
collecting acoustic signals from lathe. The acoustic signals are acquired using microphone of a new tool
and worn out tool. The acoustic signal is preconditioned, sampled, and fast Fourier transform (FFT) has
been applied to identify the frequency components to enable categorize the tool condition. The ceptrum
values are used to train Radial basis function neural network. Results show that, properly collected
acoustic signals helps achieve more than 92% accuracy of identifying the presence of frequency
components using FFT.

Keywords: fast Fourier transform (FFT), unidirectional microphone, tool wear, condition monitoring,
acoustic signal.
Introduction
In the normal machining environment, an operator knows the tool life. Once the tool tip has worn
out, he will change the tool bit or regrind the tool. This can be done by introducing automatic monitoring
system. Tool wear effects the cutting performance in the form of change in the diameter of the work
piece, deterioration of the surface finish of the work piece, increase in the cutting force, and increase in
the tool vibration. The change in diameter and increase in cutting forces are due to increase in the flank
wear land width (Vb). Monitoring of Vb is done by sensing anyone of the signals, i.e., cutting forces, tool
holder vibration, acoustic emission.
The methods used for monitoring tool wear are (I) direct and (ii) indirect. The direct methods use
measurements of volumetric loss of tool material. This procedure is done off-line. Some of the direct
methods include change in work piece dimension, optical techniques, radio-active methods and
pneumatic gauging method. The indirect methods use the measurement of cutting related parameters like
cutting forces, tool holder vibration, acoustic emission, etc.
Vibration monitoring of the machine during the machining operation is one of the important
signal measuring procedure for detecting the status of the tool wear [1].
Experimental Setup For Collecting Tool Wear Data
The experimental setup includes collection of sound from the turning machine in different turning
conditions. The sound were collected using PCB 130 series of Array Microphones coupled with ICP
sensor powered preamps and are thus referred to as ICP microphones. The 130 series provide an
extremely cost effective method for large channel count sound pressure measurements. Figures 1 ,2 and
3 shows some of the sound files associated with different turning conditions.

Figure 1: 1.wav file signal ( Refer Table 1 for turning condition)

Figure 2: 3.wav file signal ( Refer Table 1 for turning condition)

Figure 3: 4.wav file signal ( Refer Table 1 for turning condition)

Table 1: Experimental data collection for A6061

S.No Speed Feed Depth of Fx N Fy N Fz, N Flank Sound
m/min rev/min cut , mm wear wave file
Vb, name
mm

1 800 20 0.5 35 65 70 0.04 1.wav

2 800 20 1.0 40 65 75 0.15 2.wav
3 800 20 1.5 40 75 80 0.32 3.wav
4 800 30 0.5 50 75 90 0.16 4.wav
5 800 30 1.0 45 85 105 0.36 5.wav
6 800 30 1.5 60 90 110 0.05 6.wav
7 800 40 0.5 40 80 65 0.37 7.wav
8 800 40 1.0 45 80 70 0.05 8.wav
9 800 40 1.5 45 90 85 0.19 9.wav
10 1000 20 0.5 50 90 85 0.40 10.wav

Materials and Methods

Radial Basis Function

Artificial neural network (ANN) is suitable for mapping inputs and target outputs [2, 3]. The concept of
distance measure is used to associate the input and output pattern values. Radial Basis Functions (RBFs)
are capable of producing approximations to an unknown function ‘f’ from a set of input data. The
approximation is produced by passing an input point through a set of basis functions, each of which
contains one of the RBF centers, multiplying the result of each function by a coefficient and then
summing them linearly [4,5]. The data have been trained using radial basis functions. The RBF learns 27
patterns during training and testing. The radial basis function (RBF) network offers a new and more
effective method for training [6, 7]. It has the ability of fast convergence and is able to fix centers in the
hidden layer during training. Table 2 gives training and testing steps by RBF.
Table 2: Radial Basis function
Training RBF Testing

Step 1: Apply Radial Basis Function for the Step 1: Obtain cepstrum values from the
cepstrum values obtained from the vibrational data
collected vibration data Step 2: Calculate RBF as. RBF = exp ( -X )
No. of Input = 10 Step 3: Calculate. Value = F * RBF
No. of Patterns = 27 Step 4: Classify the value based on the
No. of Centre = 4 stipulated threshold to decide the status of
Calculate RBF as RBF = exp (-X) the tool wear
Calculate Matrix as G = RBF
A = GT * G
Calculate B = A-1
Calculate E = B * GT
Step 2: Calculate the Final Weight F = E * D
Step 3: Store the Final Weights in a File.

Cepstrum analysis is an important sound processing algorithm that separates high frequency noise from
low frequency information. Steps involved in converting acoustic data into features are given below:
1. Acquire acoustic data using PCB 130 series Microphone.
2. Remove zeros which does not give any information.
3. Apply linear predictive analysis.
4. Apply fast Fourier transform.
5. Apply log for the output in step 4.
6. Apply inverse fast fast Fourier transform.
7. Apply Levinson Durban equation.
8. Repeat step 3 to step 7 for every 10 samples of the data and average all values to finally get only 10
values.
9. Label the averaged patterns using the corresponding vb values.
10. Repeat step 9 for 27 different turning conditions.

Analysis of signals collected form Microphone using FFT

Single-Sided Amplitude Spectrum of y(t)
2000

1500
|Y(f)|

1000

500

0
0 500 1000 1500 2000 2500 3000 3500 4000
Frequency (Hz)

Figure :4 Frequencies versus amplitude foracoustic signal collected from Microphone

Figure 4 shows frequencies obtained for 8000 samples of a signal. The Figure shows single sided
frequencies. By analyzing the remaining samples of the signal using FFT, similar sets of frequencies are
obtained. In order to find if there is any variation on the frequencies and their amplitudes, four high
frequencies are considered from one half of the 8000 samples. These four frequencies are averaged to
find approximate frequency present in the samples.
Table 3 shows the cepstrum values for the Figures 1 to 9. Out of 27 sound wave files only 9 wave files
are shown as sample.
Table 3 Cepstrum values for the wave files given in the 9th column of table 1
Features obtained from acoustic data
Ptn 1 2 3 4 5 6 7 8 9 10
.N
o
1 1 -6.395 -2.193 2.7182 1.5176 1.4621 1.113 3.030 -1.81 - C
7.9146 I
2 1 -14.93 0.7631 11.042 1.7933 -7.41 1.237 10.42 -1.19 -
15.452
4
3 1 0.2196 -0.198 -0.054 0.3156 0.2388 -0.24 -0.303 0.140 0.2242
4 1 -3.859 -1.863 2.2723 3.7506 -0.455 -4.21 -2.077 2.111 4.2532
5 1 4.9165 -0.373 -3.835 -0.166 3.4701 1.068 -2.478 -0.40 3.6174
6 1 0.2535 -0.539 -0.713 0.1204 0.4313 0.252 0.038 -0.24 -
0.0948
7 1 -4.336 -1.732 2.7347 2.5001 -1.092 -2.49 -0.571 1.038 0.8719
8 1 0.309 -0.604 -0.780 -0.136 0.4861 0.464 0.118 -0.15 -0.196
9 1 0.9157 -0.437 -0.782 0.2436 0.620 -0.04 -0.380 -0.13 0.3291
..
27 1 0.9157 -0.437 -0.782 0.2436 0.6203 -0.04 -0.38 -0.13 0.3291 C
II
C I ( Class I) =0.1(Tool wear within the limit) C II (Class II) = 0.2 (tool worn out)

Results and discussion

The features present in the Table 3 (column 1 to 9) are used as inputs for RBF. The RBF with 3 centers
is able to correctly identify tool which is about to reach the non-acceptable wear limit. The effects of
number of center (Table 4) in identifying normal and abnormal tool wear are shown in Figures 5.

Table 4: Performance of radial basis

function centers
S.No No. of Percentage identification
centers for averaged patterns in
each category
1 1 0%
2 2 25%
3 3 75%
4 4 100 %
 2
.
target
output
1.5
Category identified

0.5

-0.5
0 1 2 3 4 5 6 7 8 9
Patterns

Figure 5 Performance of RBF when the centre is 1

Conclusions
Microphone has been used to record acoustic data during turning operation with normal condition and
completely worn out tool. Sampling was done at 44khz. The data have been converted into cepstrum
values using FFT and Levinson-Durban methods. The data have been converted into cepstrum values to
be used as inputs for RBF. The RBF with 3 centers is able to correctly identify tool which is about to
reach the non acceptable wear limit. The accuracy of tool condition classification is more than 90%.
References
[1]. Adman G Rehorn, Jin Jiang, Peter E Orban, State-of the art methods and results in tool conditioning
monitoring: a review, International Journal of Advanced Manufacturing Technology (2005) 26: 693-
710.
[2]. S.Purushothaman, Y.G.Srinivasa, (1998), A procedure for training artificial neural network with the
application of tool wear monitoring, International Journal of Production Research, Vol.36, No.3,
PP.635-651.
[3.] S.Purushothaman, Y.G.Srinivasa, (1994), A back-propagation algorithm applied to tool wear
monitoring, International Journal of M/C Tools & Manufacturing, Vol 34, No.5, pp.625-631.
[4]. S.Purushothaman, Sambasiva Rao Baragada, S. Ramakrishna, M.S.Rao, Polynomial Discriminant
Radial Basis Function for Steganalysis, International Journal of Computer Science and Security, Korea,
March 2009.
[5]. S.Purushothaman, Tool Wear Monitoring Using Artificial Neural Network Based On Extended
Kalman Filter Weight Updation With Transformed Input Patterns, Journal of Intelligent manufacturing.
[6.]Guofeng Wang, Yinhu Cui, On line tool wear monitoring based on auto associative neural network,
J. Intelligent Manufacturing, 2013,VOl.24, pp.1085-1094.
[7]. Brezak, D, Udiljak, T, Majetic, D, Novakovic B, Kasac, J, Tool wear monitoring using radial basis
function neural network, Neural Networks, 2004. Proceedings. 2004 IEEE International Joint
Conference, 25-29 July 2004, ISBN: 0-7803-8359-, INSPEC Accession Number: 8231208.