Sideband Algorithm for Automatic Wind Turbine Gearbox Fault ...
Automatic Cutting Tool Fault Detection
-
Upload
khalid-f-abdulraheem -
Category
Documents
-
view
62 -
download
5
Transcript of Automatic Cutting Tool Fault Detection
AUTOMATIC CUTTING TOOL FAULT DETECTION
USING WAVELET ANALYSIS AND AN ARTIFICIAL
NEURAL NETWORK (ANN)
Nasser Al-Mushifri, Khalid F. Abdulraheem, Waleed Abdul-Karim
ABSTRACT
In this investigation the cutting fault detection techniques have been developed based on
the wavelet analysis and artificial neural network (ANN) as automatic cutting tool fault
detection and classification. Two experimental works have been employed to capture the
vibration signals for analysis .In both experiments, the turning lathe machine is used with
high carbon steel tip cutting tool and mild steel material as work piece. In the first
experiment, the vibration measuring device (vibro-60) is used to collect and analysis the
vibration acceleration signals. The time domain statistics with different wear conditions have
been calculated using MATLAB software. The cutting condition parameters are kept constant
and the wear level is changed for both experiments. In the second experiment the Data
Acquisition system (DAQ) with LABVIEW software is used to capture the vibration signals for
cutting tool with different wear conditions. The captured vibrations data are analyzed using
continuous wavelet transform (CWT) with Morlet wavelet and Daubechies wavelet as a base
function. In general, the CWT coefficient is used to generate the inputs features to ANN for
automatic tool condition classification, with two outputs (0, 1) for healthy and (1, 0) for
faulty. The results show the effectiveness of the combed WT and ANN for automatic
classification of tool wear conditions with high success rate.
INTRODUCTION
The quality of the products is very important for customers because it is the quality that
influences the degree of satisfaction of the consumers during usage of the goods procured,
therefore the manufacturing, industrial and production unit should monitor the conditions
of machines because the production is connected to a good health of the machine. So,
condition monitoring is important for increasing machinery availability, improving
manufacturing process, productivity and reliability, and reducing maintenance costs. The
most important thing an investor is concerned is profit and how to save on the cost of
production under the market pressure and companies compotation. An efficient condition
monitoring scheme is capable of providing warnings and predicting faults at early stages.
In this work the predictive or condition-based maintenance is applied based on the wear
condition. This type is useful to monitor the condition of the cutting tool of a lathe machine
which uses a specific technique which is cutting tool vibration signal. It can be used to detect
the signal of the cutting tool faults and analyses by using the time domain statistics and
wavelet analysis as feature extraction. The extracted WT analysed parameter to be fed as
inputs features to an artificial neural network (ANN) for fault detection and classification.
The cutting fault detection techniques have been developed based on the wavelet analysis
and ANN as automatic cutting tool fault detection and classification. Two experimental
works have been employed to capture the vibration signals for analysis with different
measuring intervals. The turning lathe machine is used with high carbon steel tip cutting tool
and mild steel material as work piece. The cutting condition parameters are kept constant
(speed 280 rpm, feed rate 0.1 mm/rev, depth of cut 0.5 mm and the direction of sensor is
vertical) and the wear levels (0.0mm, 0.2mm , 0.4mm and 0.6mm) are changed for both
experiments.
Techniques for cutting tool condition monitoring can be grouped into two main categories,
direct sensing and indirect sensing techniques. D.E. Dimla Sr. (2000) has reported that direct
methods are less beneficial than indirect methods because the cutting area is largely
inaccessible, and therefore on-line monitoring cannot be carried out while the tool is
engaged in -process cutting. The direct methods use these technologies: touching trigger
probes, optical, radioactive, proximity sensors and electrical resistance measurement
techniques (I.N. Tansel 1992). Furthermore, indirect methods are widely used in cutting tool
condition monitoring and as fault detection. They are used in different techniques such as
cutting forces, acoustic emission, temperature, vibration, spindle motor current and torque
(S. S. Abuthakeer 2011). However, vibration measurement for machinery condition
monitoring is easy, less costly and yields a great deal of information that can be used to
monitor the relative motion between the tool tip and the work piece for precision of the
cutting operation (Qiao Sun 2005). Issam Abu-Mahfouz et. al. (2003) have reported that
vibration analysis is widely accepted as a tool to monitor the operating conditions of a
machine as it is nondestructive, reliable and permits continuous monitoring without
intervening with the process. This study has demonstrated tool condition monitoring
approach in turning operation based on the vibration signal collected because of the
availability, low cost, large information data, nondestructive, reliable and permits
continuous monitoring for on-line monitoring. This use is easily replaceable and very cost-
effective accelerometer.
Wavelet analysis
A short oscillating function which contains both the analysis function and the window is
known as a wavelet. Wavelets are applied to transform the signal under examination into
another representation which presents the signal information in a more amenable form
.This type of transformation is known as Wavelet Transform (WT) which is converting the
function or signal into another form to make certain features of the original signal more
useful to investigate. There are five categories of wavelet transform in TCM: time–frequency
analysis of machining signal, feature extraction, signals denoising, singularity analysis for tool
state 38 estimation, and density estimation for tool wear classification according to its multi-
resolution, sparsity and localization properties ( Zhu Kunpeng 2009). The wavelet method
overcomes the limitation of FT, by using a multi-resolution technics (time & frequency).
Hence, it has the ability to examine the signal simultaneously in time and frequency with a
flexible mathematical foundation. Time information is obtained by shifting the wavelet over
the signal. The frequencies are changed by contraction and dilatation of the wavelet
function. The wavelet analysis is more sensitive and trustable than the Fourier analysis for
recognizing the tool wear states in turning (W.Gong, 1997). Joseph Fourier in 1800 shows
the ability of superpose sines and cosines to represent other functions. The wavelets were
first mentioned by Alfred Haar in 1909. It can be moved at various locations on the signal
and also it can be squeezed to different scales.
There are some requirements for the wavelet, which are: it must have finite energy, based
on Fourier transform, the wavelet must have zero mean and for complex wavelets the
Fourier transform must be both real and vanish for negative frequencies. To make the
wavelet of chosen mother wavelet more flexible, two basic manipulations are applied which
are stretch or squeeze it (dilation) and move it (translation). The dilation of the wavelet is
governed by the dilation parameter a. The movement of the wavelet along the time axis is
governed by the translation parameter b. These shifted and dilated versions of the mother
wavelet Ψ (t) are denoted by Ψ [(t – b)/a]. The wavelet represented by (Paul S. Addison,
2002):
( )
√ (
) (1)
Where, the factor
√ is used to ensure energy preservation.
The sum over all time of the signal multiplied by scaled and shifted versions of the wavelet
function Ψ is called the continuous wavelet transform (CWT). Continuous wavelet
transforms are recognized as effective tools for both stationary and non-stationary signals.
Based on the equation (3 – 9) the CWT given as:
( ) ∫ ( ) ( )
(2)
Where, ( ) is the continuous wavelet transform (CWT), X (t) is the signal and the
superscript asterisk '*' stands for the complex conjugate.
It can be expressed in more compact form as an inner product:
⟨ ⟩ (3)
The windowing techniques with variable-size regions in wavelet analysis can be overcome
the limitations of the STFT. Wavelet analysis allows the use of shorter time intervals where
more precise high frequency information is desirable and long regions for low frequency
information. Sometimes the wavelet is irregular and asymmetric waveform of effectively
limited duration (average value zero), so the varieties of wavelets (Wavelet Families) are
exist, and an analyst can choose from the wavelet families that suits his application best.
This project focuses on Daubechies wavelets (db10) and Morlet wavelet (morl) as basis
function based on the role of features extraction, because they are more similar
characteristics to the extracted signals. The morl wavelet is known by the following Equation
(Semmlow, 2004). Figure 1 shows the morl wavelet.
( ) ( √
) (4)
Figure 1 Morlet wavelet.
The names of Daubechies family wavelets are signed db N (N is the order) as Figure 2 shows
the db 10 wavelet. The Daubechies wavelet is defined as given in equation (5) (Daniel T.L.
Lee, 2004).
( ) ∑ √ ( ) (5)
Where, Φ(t) is scaling function and = (-1) k α – k +1 , If N= 1, then α0= α = 1
Figure 2 Daubechies wavelet db 10
ARTIFICIAL NEURAL NETWORK (ANN)
The idea of Artificial Neural Network (ANN) is, creating a computing system that simulates
the biological neural systems of the human brain. The artificial neural network is particularly
useful in the modeling of nonlinear mapping, and also in the recognition of distinctive
features from chaotic input data even if it is not complete. The behavior of ANN modifies in
response to its environment. The inputs will self-adjust while a set of them are given to the
network to produce consistent responses through a process called learning. The learning
process can change the weights systematically in order to achieve some desired results for a
given set of inputs. The types of learning are, supervised and unsupervised; the supervised
has been selected based on the environment knowing. The popular algorithm related to the
supervised learning is known as the Back-Propagation.
The construction of ANN involves the determination of the network properties depending on
the network topology (connectivity), the type of connections, the order of connections, and
the weight range. Moreover, it determines the node properties like the activation range and
the activation function. Also, in dynamic system the ANN determines weight initialization
scheme, the activation calculating formula and the learning rule. A large number of
researchers presented application of neural network models in Tool Condition Monitoring
(TCM) and classification of tool wear. Artificial neural network (ANN) is useful as online
prediction of tool wear based on back propagation network (R R Srikant, 2011). The multi-
layer feed-forward neural network with a back propagation (FFBP) training algorithm is
successful in TCM as tool fault detection and classification (Issam Abu-Mahfouz 2003). In this
project the ANN is used as fault detaction and wear condition classification based on Multi-
layer feed-forward with back propagation.
To classify and quantify the fault of cutting tool, neural network is used with signal analysis
techniques. Neural network method for automatically classifying the machine condition
from the vibration time series have used by McCormick and Nandi (1979). Artificial neural
network contain many connected neurons, which work as receiver for the impulses from
input or other neurons. These neurons transform the input by giving the outcome to the
output or other neurons. Also, ANN consists of different layers of connected neurons, which
receive the input from the previous layer and transfer the output to the succeeding layer.
Figure 3 shows the model of a neuron, where the inputs are forwarded to the neuron and
multiplied by their synaptic weights. Then, the outcome is forwarded to sum in summing
junction and it is activated by the activation function. The inputs of the activation function
are affected by the bias (bi), so it will increase if positive and decrease in the case of
negative. Finally, the output will be given. The learning and storing of the knowledge will be
possible by this model of the ANN.
Figure 3 The model of a neuron
Typically, in neural network architectures two types of layers are organized in the shape of a
layered neural network. There are the signal-layer feed-forward perceptron (SLP) neural
network and the feed-forward multilayer perceptron (MLP) ANN. The arrangement of
neurons in each of the layers is entirely dependent on the user, hence they have the ability
to represent a large range of output and input patterns. The Figure 4 shows the limitations
in the range of functions or processes that they can represent in signal-layer feed-forward
(SLP) neural network. However, the feed-forward multilayer perceptron (MLP) neural
network is selected in this study; because it has a wide range of processes and more
powerful representation capacity which can be achieved by using more than one layer, as
shown in Figure 4 .
Figure 4 Typical diagram of a single-layer perceptron
Figure 5 Typical diagram of multi-layered feed-forward ANN
EXPERIMENTAL SETUP
In This study the experimental work demonstrates the tool condition monitoring approach in
turning operation based on the vibration monitoring. This experiment was conducted in the
Workshop in Caledonian College of Engineering (CCE) using (TM-35) lathe machine.
However, two experiments have been applied to monitor the tool wear based on vibration
signals collected from two different measuring devices. These experiments work with same
procedure and parameters of the cutting conditions, which are speed 280 rpm, feed rate 0.1
mm/rev, depth of cut 0.5 mm and the direction of sensor is vertical. Moreover, all these
experiments have been setup based on the different wear conditions, which are healthy
tool, 0.2mm, 0.4mm and 0.6mm wear levels.
Preparing For the Experiments
This experimental work has a small preparation which can create the wear levels at the tip of
carbide inserts. In this experiment first, the carbide inserts is fixed on the tool holder and the
grinding wheel is gagged at the lathe machine as shown in Figure 6 (a). Then, the lathe
machine has to be adjusted to make wear at the cutting tool using the grinding wheel with
speed of 450 rpm. When the lathe machine is adjusted to a depth of cut 0.2 mm the wear is
created about 0.2mm wear level. Finally, the wear at carbide insert tips is measured using
the microscope. The same procedure is used to create the 0.4mm and 0.6 wear level. Figure
6 shows an example of the process of creating wear levels.
Figure 6 An example of the process of creating wear levels, (a) use of grading wheel, (b) the
wear created in the tip of carbide insert and (b) the microscope.
Experiment ( I )
Experiment (I) is applied to measure the vibration of the cutting operation using Predictive
Machine VIBROTEST 60 to handle the vibration analyzer and data collector. The vibration
signals are measured using vibro (60) device for different wear conditions, in which health
tool, 0.2 mm 0.4 and 0.6 mm wears. In this experiment the workpiece and cutting tool is
fixed at the lathe machine and the distance between the work piece and the cutting tool is
adjusted. First, the new cutting tool with (0.0mm) wear is employed with the work piece
and the direction of the accelerometer sensor (coated by aluminum coil for safety) is vertical
as show in the Figure 7.
Figure 7 The adjusting workpiece and cutting tool at the lathe machine
Also, the acceleration sensor has to be connected between the VIBROTEST 60 and the
cutting tool holder using AC-437 Cable to collect the signals. While the lathe machine
rotates and the cutting process starts, the acceleration sensor sends the signals to
VIBROTEST 60. After that, the data will be saved in PCMCIA card, which can be transferred to
a PC with a card reader and XMS program. Finally, the selective data is analyzed using
MATLAB. The same procedure is applied to the other cutting tools with 0.2 mm, 0.4 mm and
0.6 mm wear levels. Figure 8 shows the experiment (I) setup.
Figure 8 The experimental (I) setup
Experiment (II)
In this experiment the vibration signals are captured by data acquisition card from National
Instruments (DAQ Card) with LABVIEW software. However, the cutting condition is stable for
both experiments, but with different accelerometer and connecting cables.
The experiment (II) setup is same as the experiment (I), but this experiment is directly
connected to a PC with a data acquisition card (DAQ card) (Type SHC 68-68-EPM) and
LABVIEW software and this experimental works with 16000 sampling rate. In this experiment
the vibration signals of the cutting tool with different wear conditions (healthy, 0.2mm,
0.4mm and 0.6mm) are captured by the accelerometer sensor to shift and convert these
signals from analogue to digital form using a PC with data acquisition card from National
Instruments (DAQ Card) and LABVIEW software. Figure 9 shows the experimental (II) setup.
Figure 9 the experimental (II) setup
Finally, the LABVIEW shows the features of the vibration signals as time domain and
frequency domain. Also, for further analyzing the suitable vibration signals the MATLAB is
used to analyses the signals based on Wavelet analysis and ANN as automatic cutting tool
fault detection and classification.
THE EXPERIMENTAL RESULTS AND DISCUSSIONS
This section presents the experimental results and discussion of the test vibration measuring
device (vibro 60) and based on acceleration signals and time domain statistics. Also, it shows
the results and discussion of Data Acquisition system (DAQ) with LABVIEW software using
(CWT). Finally, it illustrates Automatic fault detection and classification of cutting tool using
ANN.
PART I: USE THE TEST VIBRATION MEASURING DEVICE (VIBRO 60)
The vibration measuring device (vibro-60) is used to select the best vibration signals to show
the results based on the acceleration signals and time domain statistics with different wear
condition using MATLAB software.
Acceleration signals
Figure 10 shows results show the different vibration signals acceleration from vibro 60 for
new tool (healthy) and different wears at speed 280 rpm, feed rate 0.1 mm/rev, depth of cut
0.5mm and the direction of sensor is vertical. These accelerations show the ability to
recognize between the healthy condition and faulty condition of the cutting tool. All the given
graphs are plotted based on time duration of 17 sec.
( a ) ( b )
( c ) ( d )
Figure 10 Acceleration form for (a) new tool, (b) 0.2 mm wear,(c) 0.4 mm wear and (d) 0.6
mm wear.
Time domain statistics
0 2 4 6 8 10 12 14 160
1
2
3
4
Time(sec)
Accela
rtaio
n(g)
New
tool
0 2 4 6 8 10 12 14 160
1
2
3
4
Time(sec)
Accela
ratio
n(g)
wear
0.2
0 2 4 6 8 10 12 14 160
1
2
3
4
Time(sec)
Accela
ratio
n(g)
wear
0.4
0 2 4 6 8 10 12 14 160
1
2
3
4
Time(sec)
Accela
ratio
n(g)
wear
0.6
The following results are demonstrated a comparative graphs of time domain statistics for
healthy tool and different wears condition. Time domain parameters show different features
of the vibration signal to recognize between the healthy condition and different wears
condition of the cutting tool. Figure 11 shows the effectiveness of the parameters of time
domain namely (Peak, RMS, Crest Factor, Kurtosis, Impulse Factor and Shape Factor) on the
wear conditions.
( a ) ( b )
( c ) ( d )
( e ) ( f )
Figure 11 (a-f) Time domain features for deferent wear condition: (a) Peak, (b) RMS, (c) Crest
Factor, (d) Kurtosis, (e) Impulse Factor and (f) Shape Factor.
0 0.2 0.4 0.60
0.1
0.2
0.3
0.4
0.5
0.6
Tool wear (mm)
Pe
ak
0 0.2 0.4 0.60
0.05
0.1
0.15
0.2
0.25
Tool wear (mm)R
MS
0 0.2 0.4 0.6
2.6
2.8
3
3.2
Tool wear (mm)
Crest factor
0 0.2 0.4 0.60
1
2
3
4
5
6
7
Tool wear (mm)
Ku
rto
sis
0 0.2 0.4 0.60
0.5
1
1.5
2
Tool wear (mm)
Im
pu
lse
fa
cto
r
0 0.2 0.4 0.60
0.2
0.4
0.6
0.8
Tool wear (mm)
Sh
ap
e fa
cto
r
The peak level is generally increasing as the wear of the tool increase Figure 11 (a). This
reflects the increasing in the overall vibration level as shown in Figure 10.The RMS level
(root-mean-square) for wear condition Figure 11 (b) shows that the RMS values increase as
the wear of the tool increase from healthy to 0.6 mm wear. This because the average
vibration amplitude (mean) and the effect of spurious peaks increase by developing the tool
wears.
Figures 11 (c, d) shows the crest factor and the kurtosis factor decrease as the tool wear
increase. They are decreasing with progressive failure of the cutting tool because the RMS
level generally increases. Moreover, the crest factor expresses that the instrument has the
capability of correct measurement how much distorted waveform. And, kurtosis is any
measure of the "peakedness" of the probability distribution of a real-valued random
variable. So, the kurtosis is high for healthy tool because the signal sharpness is large (sharp
tool) and then signals became more noisy (i.e. the values distributed over wide raise) as the
wear increase.
The impulse factor and shape factor for cutting tool condition Figure 11 (e, f) show that as
the tool wear increase the values of impulse factor and the shape factor increase. However,
the impulse factor is referring to control the disorder of the data. And, the shape factor is
referring to how much the tool will deform when pressure is applied to it. So, the shape
factor is increasing during the wear increasing.
PART II: USING DATA ACQUISITION SYSTEM (DAQ) WITH LABVIEW SOFTWARE
This section presents the results of applying the Continuous Wavelet Transform (CWT) for
cutting tool with different wear condition. The CWT is used also as features extraction
method for generate the inputs feature to ANN. Morlet wavelet (morl) and Daubechies
Wavelets (db 10) has been used as a mother wavelet function while obtains the CWT
coefficients.
Continuous Wavelet Transform (CWT) Analysis
Figure 12 shows the results of wavelet analysis for different vibration signals captured by
DAQ for new tool (healthy) and different wears at speed 280 rpm, feed rate 0.1 mm/rev,
depth of cut 0.5mm and the direction of sensor is vertical. In Figure 12 the results are
demonstrated comparative graphs of kurtosis factor of 20 wavelet coefficients for healthy
tool and (0.2, 0.4, 0.6) mm wear, respectively. These graphs present the effectiveness of
kurtosis factor in wavelet scale, and then compare it to histograms of healthy tool and each
different wears. This technique shows the ability to recognize between the healthy condition
and faulty condition.
Figure 12 The Kurtosis distribution for wavelet transform scales [healthy tool and (0.2, 0.4,
0.6) mm wear]
The Kurtosis distribution of wavelet transform scales Figure 12 presents that the kurtosis of
healthy tool is higher than the fault condition. As the sharp tool produces a signal with less
randomness and as the wear progress the randomness of the signal is increased that
produce a flat signal distribution as a result the kurtosis value will decrease. This clear in the
histograms for healthy tool and (0.2, 0.4, 0.6) mm wears shown in Figure 13 (a-e).
Figure 13 (a) The wavelet kurtosis by the wear condition and (b-e) the histograms of healthy
tool and (0.2, 0.4, 0.6) mm wear.
0 2 4 6 8 10 12 14 16 18 200
10
20
30
40
50
60
scale
wavele
t kurtosis
Healthy
0.2 wear
0.4 wear
0.6 wear
To change the wavelet scale to frequency (Hz), this equation is applied: Fa = Fc / a . ∆ Where,
a is a scale, ∆ is the sampling period (0.1), Fc is the center frequency of a wavelet in Hz and Fa
is the pseudo-frequency corresponding to the scale a in Hz. In the selected signal was in 280
rpm, so the frequency equal to (280/60) 4.6 Hz. Figure 14 shows the kurtosis distribution of
wavelet transform at frequency (Hz) for healthy tool and (0.2, 0.4, 0.6) mm wear,
respectively.
Figure 14 The Kurtosis distribution for wavelet transform at frequency (Hz) [healthy tool and
(0.2, 0.4, 0.6) mm wear]
The wavelet kurtosis decrease when tool wear increase as shown in above Figure 14 this is
corresponding to frequency in hertz (
). That is referring to the damping of the cutting
tool increase as result of increasing in the wear. When the wear increase the contact area of
the cutting tool with the workpiece also increase. That makes the friction increase, so the
damping increase and the machine frequency decrease for lately the peak at 4.06 Hz for
healthy tool and shifted to 1.35 Hz for 0.2 mm wear, again decreased for 0.4 and 0.6 mm
wear. All that is presented in this equation: √ where, ωd is damped
frequency, ωn is the undamped angular frequency and ζ is damping ratio which increase by
increase tool wear.
AUTOMATIC FAULT DETECTION AND CLASSIFICATION OF CUTTING TOOL USING ANN
Automatic fault detection and classification of cutting tool condition using the features of
wavelet and Artificial Neural Network (ANN) model is proposed for this project. By using The
Artificial Neural Network (ANN) to classify the tool wear conditions the model of ANN is
created based on feed-forward Multi-Layer Perceptron (MLP) and Back Propagation. The
result features (peak, RMS, crest factor, kurtosis, shape factor and impulse factor) healthy
condition and wear condition are feed to ANN to classify the wear condition. The signal
consists of 38400 data for each condition (wear & healthy) and then 10 coefficients are
taken for each of these wear conditions. To build the ANN model six features are extracted
from 10 coefficients for each condition, then the values divided into 30 (5x6) values for
0 1 2 3 4 5 6 7 8 90
10
20
30
40
50
60
frequency (Hz)
wa
ve
let
ku
rto
sis
Healthy
0.2 wear
0.4 wear
0.6 wear
training and 30 values (5x6) for testing as shows in Table 3 and Table 4, respectively. The
healthy condition is normalized as (0 1) and wear condition as (1 0) for training targets.
Table 3 Training values feed to ANN
Training
Condition RMS Kurtosis Peak Crest Impulse Shape Target
Healthy
0.4668 11.1281 3.8155 8.1746 1.5383 0.1882 0 1
0.4041 7.2317 2.6389 6.5300 0.8989 0.1377 0 1
0.3559 7.6381 2.5896 7.2761 0.9182 0.1262 0 1
0.3279 3.6604 1.6874 5.1461 0.5828 0.1132 0 1
0.4024 6.3357 2.4941 6.1978 0.9593 0.1548 0 1
Wear
2.0584 4.6147 11.3996 5.538 18.0085 3.2518 1 0
2.3546 3.4431 9.4807 4.0266 17.4506 4.3339 1 0
2.3849 4.6438 14.0839 5.9054 26.0475 4.4108 1 0
2.1151 3.3395 8.3399 3.943 13.8874 3.522 1 0
2.1816 4.8249 13.9964 6.4157 23.9821 3.738 1 0
Table 4 Testing values feed to ANN
Testing
Condition RMS Kurtosis Peak Crest Impulse Shape
Healthy
0.4134 12.8182 4.0192 9.7217 1.4858 0.1528
0.3218 3.9022 1.4742 4.5816 0.4926 0.1075
0.3162 3.9536 1.5259 4.826 0.4944 0.1024
0.3274 4.339 1.6683 5.0959 0.5459 0.1071
0.2901 3.5146 1.4233 4.9061 0.4446 0.0906
Wear
2.7586 18.9119 27.5982 10.0046 52.6953 5.2671
2.4656 5.026 15.5254 6.2967 29.5224 4.6885
2.6058 9.7085 25.6753 9.8532 50.4149 5.1166
2.3226 3.511 8.6374 3.7189 15.7201 4.2271
2.6525 8.3905 21.069 7.9431 42.2863 5.3237
The ANN model is created using input layer with six nodes (extracted features), two hidden
layers consist five nodes for each and output layer as shown in Figure 15. Back Propagation
is applied to minimize the Mean Square Error (MSE) between the ANN outputs and the
desired target values.
Figure 15 The model of ANN
In this model two stages are applied which are training stage and testing stage. It is trained
with 10E-10 training goal (MSE), 0.52044 training rate, with six attribute (features) and the
maximum No. of iteration (epochs) of 1000 are selected. Figure 16 shows the result of
training process, in which it reached the desired goal stopping criteria after 27 epochs.
Figure 16 Training process of ANN
Figure 17 presents the ANN parameters and the training stage performance in which the
ANN performance is shown as 98% success rate.
Figure 17 ANN parameters and the training stage performance
The regression curve for both training targets and the ANN output is shown in Figure 18, a
good correlation between the both can be concluded. The results for ANN classification for
tool wear condition shown success rate based on the given six features.
Figure 18 The regression curve for both training targets and the ANN output
CONCLUSIONS
Based on the obtained results the overall conclusion can be summarized as follow:
The acceleration of vibration signals and time domain statistic shows the ability to
recognize the tool wear condition.
Kurtosis Factor proved to be more accurate fault detection parameter comparing to
other wavelet parameters (Peak, RMS, crest factor, Impulse Factor and shape factor).
For more accurate fault detection of the cutting tool a new techniques has been
developed, which are wavelet kurtosis factor and the histograms throughout using
Morlet wavelet and Daubechies Wavelets as a mother wavelet function (similarity with
the extracted fault pulses shape). This technique shows the ability to recognize between
the healthy and wear conditions.
The wavelet analysis is selected for cutting tool vibration signal features extraction. The
advantage of wavelet analysis is proven as a multi resolution, scaling and shifting of the
wavelet through the vibrational signal.
For high performance of the extracted wavelet features; the features are normalized
between 0 and 1 in order to be the inputs in ANN.
The ANN model based on supervised learning capability of Multi-Layer Perceptron (MLP)
and Back Propagation has shown effectiveness to be as automatic cutting tool fault
detection and classification, as proven that the training process has been reached the
desired goal stopping criteria after 27 epochs. And the ANN performance is shown as
98% success rate.
This project is proved the successful correlation between the wavelet transform (WT)
and the tool wear condition based on the obtained results.
References:
A.C.McCormick and A.K.Nandi (1997): Classification of the rotating machine condition using
artificial neural networks, Proceedings Instrumentation Mechanical Engineers, Vol.211, pp.
439-450.
D.E. Dimla Sr. and P.M. Lister (2000): On-line metal cutting tool condition monitoring force
and vibration analyses, International Journal of Machine Tools & Manufacture, vol. 40, pp.
739–768.
Daniel T. L. Lee and Akio Yamamoto (1994): Wavelet Analysis: Theory and Applications,
Hewlett-Packard Journal, vol. 45, Issue 6.
I.N. Tansel, O. Rodriguez and C. Mekdeci (1992): Detection of tool breakage in micro-drilling
operation with RCE neural networks, ASME PED Engng Syst. Design Analysis, vol. 47, pp. 83–
88.
Issam Abu-Mahfouz (2003): Drilling wear detection and classification using vibration signals
and artificial neural network, International Journal of Machine Tools & Manufacture, vol. 43,
pp. 707–720.
Paul S Addison (2002): The Illustrated Wavelet Transform Handbook: Introductory Theory
and Applications in Science, Engineering, Medicine and Finance, Institute of Physics
Publishing, Bristol and Philadelphia.
Qiao Sun, Ying Tang, Wei Yang Lu and Yuan Ji (2005): Feature Extraction with Discrete
Wavelet Transform for Drill Wear Monitoring, Journal of Vibration and Control, vol. 11, pp.
1375–1396.
R. R. Srikant, P. Vamsi Krishna and N. D. Rao (2011): Online tool wear prediction in wet
machining using modified back propagation neural network, Journal of Engineering
Manufacture, Part B, vol.225, pp.1009- 1018.
Semmlow J.L.( 2004): Bio signal and biomedical image processing, Marcel Dekker, Inc. New
York, USA, ISBN0-8247-4803-4
W. Gong, T. Obikawa, T. Shirakashi (1997): Monitoring of tool wear states in turning based
on wavelet analysis, JSME International Journal (Series C), vol. 40, pp. 447–453.
Zhu Kunpeng , Wong Yoke San and Hong Geok Soon (2009): Wavelet analysis of sensor
signals for tool condition monitoring: A review and some new results, International Journal
of Machine Tools & Manufacture , vol. 49, pp. 537–553.