Engineering Applications of Arti®cial Intelligence 13 (2000) 249±261

www.elsevier.com/locate/engappai

Multi-sensor integration for on-line tool wear estimation through

arti®cial neural networks and fuzzy neural network
R.J. Kuo*
Department of Industrial Engineering, National Taipei University of Technology, Taipei 106, Taiwan
Received 1 March 1999; accepted 1 January 2000

Abstract

On-line tool wear estimation plays a critical role in industrial automation for higher productivity and product quality. In
addition, an appropriate and timely decision for tool change is required in machining systems. Thus, this paper develops an
estimation system through integration of two promising technologies, arti®cial neural networks (ANNs) and fuzzy logic. The
proposed system consists of ®ve components: (1) data collection, (2) feature extraction, (3) pattern recognition, (4) multi-sensor
integration, and (5) tool/work distance compensation. Two di€erent networks, a feedforward neural network with an error
backpropagation learning algorithm and a counterpropagation neural network, are employed to recognize the extracted features
and provide a comparison of these two networks based on accuracy and speed. Meanwhile, in order to enhance the accuracy of
the estimation result, this research work applies multiple sensors for detection. The data from multiple sensors are integrated
through the proposed fuzzy logic model. Such a model is self-organizing and self-adjusting, learning from experience. Physical
experiments of the metal cutting process are implemented to evaluate the proposed system. The results showed that the
proposed system can signi®cantly increase the accuracy of the product pro®le when compared to the conventional approaches,
like multiple regression and a single ANN. 7 2000 Elsevier Science Ltd. All rights reserved.

Keywords: Production and process control; Metal cutting; Multi-sensor integration; Fuzzy neural networks; Arti®cial neural networks; Fuzzy
logic

1. Introduction Arti®cial neural networks (ANNs), fuzzy logic, and

genetic systems constitute three independent research
Wear of the cutting tool is an inevitable result of the ®elds regarding sixth generation systems (SGS). Motiv-
metal cutting process. Since undesirable e€ects of tool ated by the results in each of these areas and the po-
wear include: (1) a loss in the dimensional accuracy of tential for mutual progress in computational modeling,
the ®nished product and (2) possible damage to the an integration of these concepts is very important
workpiece, the on-line prediction of cutting tool wear (Kosko, 1992). ANNs and the fuzzy model have been
becomes crucial. To date, it remains one of the major used in many application areas (Lippmann 1987; Lee
obstacles in the optimization of the metal cutting pro- 1990a, 1990b), and each pairing has its own advan-
cess and in the full implementation of unmanned tages and disadvantages. Therefore, how to success-
machining. It is especially important for precision ¯ex- fully integrate these two approaches, ANNs and fuzzy
ible manufacturing systems (PFMS). Thus, developing modeling, for the machining systems is the main focus
an intelligent estimation system for tool wear is im- of this research.
portant. This research compares two di€erent neural net-
works, a feedforward network with an error backpro-
pagation (EBP) learning algorithm and a counter-
* Tel.: +886-2-2771-2171; fax: +886-2-2731-7168. propagation network, for pattern recognition. More-
E-mail address: [email protected] (R.J. Kuo). over, a fuzzy neural network (FNN) is proposed which

0952-1976/00/$ - see front matter 7 2000 Elsevier Science Ltd. All rights reserved.
PII: S 0 9 5 2 - 1 9 7 6 ( 0 0 ) 0 0 0 0 8 - 7
250 R.J. Kuo / Engineering Applications of Arti®cial Intelligence 13 (2000) 249±261

Nomenclature

e the sample number m the number of the categories of each input

meij the ijth membership function value for sample variable
e ze the desired output for sample e
sij the scalar of the ijth membership function fe‰ j Š the ®ring strength of rule [ j ] for sample e
cij the center of the ijth membership function a‰ j Š, i the coecient of input x i
sij the variance or width of the ijth membership b‰ j Š the coecient of input rule j
function x ei the ith input of sample e
we‰ j Š the intersection of the membership function E the cost function
values for rule [ j ], sample e Oe the actual output of sample e
n the number of the input variables Z the training rate

is self-organizing and self-adjusting, and is able to of ANNs and fuzzy logic and their applications in tool
learn from experience. For the FNN, the inputs and wear monitoring will be discussed.
outputs are partitioned by Kohonen's feature mapping
and the premise and consequence parameters are 2.1. ANNs in fuzzy modeling
updated via an EBP-type learning algorithm.
Physical experiments will be used in order to evalu- Generally, the traditional fuzzy system, as men-
ate the proposed system. Three di€erent kinds of sen- tioned above, is based on the experts' knowledge, how-
sors (force, vibration, and acoustic emission) are used ever, it is very subjective. Besides, it is very dicult to
to detect sensory information about the cutting tool acquire sound knowledge and ®nd the required exper-
wear. The basic structure of the estimation system is tise (Jang, 1992). Recently, the ANN learning algor-
to acquire the sensory signal pattern from each sensor ithm has been applied to improve the performance of
®rst and then extract the features from the sensory sig- a fuzzy system and shown to be a new and promising
nal pattern. ANNs are employed to recognize the fea- approach. Takagi and Hayashi (1991) introduced a
tures from each sensor. Each ANN provides a feedforward ANN into fuzzy inference. Each rule is
decision, the amount of tool wear, for each sensor. represented by an ANN, while all the membership
Then the multi-sensor integration method (FNN) is functions are represented by only one ANN. Jang
applied to combine all the decisions from di€erent sen- (1991, 1992) and Jang and Sun (1993) proposed a
sors. The proposed method is able to predict the con- method which transforms the fuzzy inference system
tinuous amount of tool wear better than multiple into a functional equivalent adaptive network, and
regression and ANN both in terms of speed and accu- then employ the EBP-type algorithm to update the
racy. In addition, the predicted amount of tool wear is premise parameters and a least squares method to
used to adjust the distance between the working ma- identify the consequence parameters. Meanwhile,
terial and cutting tool in order to increase the pre- Wang and Mendal (1992), Shibata et al. (1992) and
cision of the ®nished product. Fukuda and Shibata (1992) also presented similar
The rest of this paper is organized as follows. In methods. Nakayama et al. (1992) proposed a so-called
Section 2, the literature survey is presented, while Sec- FNN which has a special structure for realizing a
tion 3 presents the methodology which discusses the fuzzy inference system. Each membership function
proposed estimation system. Section 4 explains how consists of one or two sigmoid functions for each
the experiments are performed. The data obtained inference rule. Lin and Lee (1991, 1994) proposed the
from the experiments described in Section 4 are used so-called neural-network-based fuzzy logic control sys-
to evaluate the proposed system and the results and tem (NN-FLCS). They introduced the low-level learn-
discussions are given in Section 5. Finally, Section 6 ing power of neural networks in the fuzzy logic system
presents the conclusions. and provided high-level human-understandable mean-
ing to the normal connectionist architecture. Recently,
Kuo and Xue (1999a, 1999b) proposed a novel fuzzy
neural network whose inputs, outputs, and weights are
2. Background all non-symmetrical Gaussian functions. The learning
algorithm is a EBP-type learning prodedure. Further-
ANNS and fuzzy logic have been widely applied in more, the EBP-type learning algorithm is integrated
many areas (Hertz et al., 1992; Freeman and Skapura, with a genetic algorithm to speed up the training time
1991; Kosko, 1992). In the following, the integration and to avoid local minima (Kuo et al., 2000).
 R.J. Kuo / Engineering Applications of Arti®cial Intelligence 13 (2000) 249±261 251

2.2. Tool wear monitoring through ANNs and fuzzy 94 and 92% accuracy for classi®cation into two and
logic three levels of tool wear, respectively. Tansel (1992)
used ART2 to combine the information from the
There has already been much research done applying dynamometer and laser vibrometer in drilling. The
ANNs in the area of machining. Tansel (1990) devel- proposed system accurately detects the pre-failure
oped two ANN systems to represent cutting dynamics. phase for all of the cases.
Then, Tansel and Laughlin (1991) used adaptive reson- Recently, fuzzy models have also been employed in
ance theory ART2 for the detection of tool breakage tool wear monitoring where input is always divided
in milling operations, which provided a 97.2% success into several groups and the boundary between these
rate. Guillot and Oua® (1991) provided time domain groups is vague, It is very similar to tool wear whose
inputs to a feedforward, three-layer ANN which ident- status is fuzzy, especially when multiple sensors are
i®ed tool breakage at its output for milling. Similar ap- used. It has been shown how to use fuzzy models to
plications can also be found, e.g., Malakooti and recognize the fuzziness of tool wear status (Kuo,
Zhou (1992), Khanchustambham and Zhang (1992), 1994). Similarly, the tool wear monitored is the tool
Elanayar and Shin (1992) and Dan et al. (1996). How- wear state instead of the continuous value.
ever, all of these e€orts consider the case of a single
sensor only.
Rangwala and Dornfeld (1987, 1989, 1990) and 3. Methodology
Rangwala (1988) applied ANNs for monitoring tool
wear states in a turning operation. A multiple sensor In this section, a system is proposed, which can pre-
scheme utilizing cutting force and acoustic emission in- dict the amount of tool wear on-line. The system con-
formation was presented. The results showed a 95% sists of ®ve components: (1) data acquisition, (2)
success rate for classifying binary tool wear states, feature extraction, (3) pattern recognition, (4) multi-
fresh and worn. sensor integration, and (5) tool/work distance compen-
Chryssolouris and Domroese (1988, 1989) proposed sation. Fig. 1 shows the general structure of the esti-
an intelligent controller which uses a multi-sensor mation system. Each component is discussed in the
approach for process monitoring. The study focuses following subsections.
on the module which integrates the sensor-based infor-
mation to provide the controller with the best possible 3.1. Data acquistion
estimates for the tool wear and wear rate. Tests indi-
cated that when compared to the group method of This system ®rst collects a sensory signal pattern,
data handling GMDH and least-squares regression which corresponds to particular characteristics of the
techniques, ANNs were more e€ective at learning a re- process. In Fig. 1, it is assumed that there are three
lationship for providing parameters and estimates, es- sensors used. From these three sensors, three sensory
pecially when the relationship between the sensor- signal patterns can be collected through the acquisition
based information and the actual parameter is non- system.
linear. Thereafter, a statistical approach (Chryssolouris
et al., 1991, 1992) was used. Before applying the above 3.2. Feature extraction
mentioned statistical approach for integration, the con-
®dence distance measure for the support of sensor i by However, di€erent sensory signal patterns or signa-
sensor j, de®ned as dij ˆ 2A, where A is the area under tures correspond to the state of the tool as it wears. In
the probability density curve Pi yjx i † between x i and order to analyze the relationship between the signal
x j , is used to eliminate the non-consensus sensory patterns and the amounts of tool wear, it is critical to
values ®rst. extract the features of the sensory signal patterns in
Masory (1991) proposed a tool wear model based advance. Two di€erent kinds of feature extraction
on the EBP learning algorithm of ANNs. Though this methods, time series analyzer and frequency analyzer,
research tried to predict the continuous amount of are used because of the sensitivity of the signals.
tool wear, only a single cutting condition was con- Intuitively, the sensory signal patterns obtained
sidered. Similarly, ANNs were also applied by from sensors are a time series. By analyzing the time
Kamarthi et al. (1991) as the pattern recognizer, while series data, a simple mathematical time series model
the input vector was the parameter of ARMA(8, 8) can be constructed. This model is able to represent the
and the network used was a Kohonen's feature map- time series data in a concise way. Then, the coecients
ping. Force and vibration sensors were used in this ap- of the model can be treated as the features. There are
plication. Leem and Dreyfus (1992) also applied three di€erent kinds of models that can be used. Thus,
Kohonen's feature map for sensor fusion in turning. it is necessary to determine which model, autoregres-
The results showed that the proposed network achieves sive model (AR), moving average model (MA), or
252 R.J. Kuo / Engineering Applications of Arti®cial Intelligence 13 (2000) 249±261

ARMA, is more suitable for the current data. For fre- and applied to practical problems. Each poses its own
quency analysis, the Fourier transformation (FT) is advantages and disadvantages. Thus, this research will
used to transform the time series data into the fre- try to test and compare two ANNs, a feedforward net-
quency domain. The algorithm used is the FFT. Five work with an error backpropagation (EBP) learning
peak values of the frequency data points are selected algorithm and a counterpropagation network (CPN)
as the features. (Hertz et al., 1992). For the feedforward network, the
EBP learning algorithm is applied to recognize the fea-
3.3. Pattern recognition tures extracted from the above section, feature extrac-
tion, since it is easy to understand and implement.
In this component, ANNs are employed to recognize Due to the slow training speed of the EBP learning al-
the features corresponding to di€erent amounts of tool gorithm, three fuzzy decision tables developed by Kuo
wear. The inputs of the network are from the above (1995) are employed to dynamically adjust three train-
feature extraction algorithms. Thus, if the number of ing parameters (training rate, momentum, and steep-
features is n, then the number of input nodes is also n. ness of activation function).
The number of outputs, which is the amount of tool It should be reiterated that a single network is used
wear, is one. for each sensor. Thus, if there are n sensors, then n
There have already been several ANNs developed total ANNs should be implemented. Through these

Fig. 1. Multi-sensor integration system.

 R.J. Kuo / Engineering Applications of Arti®cial Intelligence 13 (2000) 249±261 253

networks, n di€erent predictions can be obtained. parameters, and A is the linguistic term. As mentioned,
Next, a new approach for multi-sensor integration, the shape and position of the membership function
FNN, will be employed in order to integrate these pre- will change when any of s, c, or s are changed. Every
dicted amounts of tool wear. node in layer 3 possesses the capability of multipli-
cation. It is equivalent to the meaning of the ®ring
3.4. Multi-sensor integration strength in fuzzy modeling. Layer 4 calculates the ith
®ring strength proportional to the sum of all the ®ring
In this component, a new multi-sensor integration strengths, while layer 5 is the output layer which com-
method (FNN) is employed. Basically, FNN consists bines all the control action values from all the infer-
of two stages as follows. ence rules.
The weights connecting any two nodes are 1 except
3.4.1. Self-organizing stage (SOS) the weights between layers 4 and 5. They are the con-
In this stage, the membership function and the con- sequences for each rule. Since Takagi's model is
sequence are de®ned. The procedures of SOS are as employed, these consequences are regression models
follows. which are denoted as fi ,
X !
3.4.1.1. Clustering. Kohonen's feature mapping is used fi ˆ ai, j x j ‡ bi 2†
to divide the training data into a speci®ed number of j
groups. Based on the results of Kohonen's feature
mapping, each group corresponds to an inference rule. where j is number of input, aij is the coecient, x is
the independent variable or predicted amount of tool
3.4.1.2. Shape and position determination. Di€erent wear, and bi is the constant term. The proposed learn-
membership functions can be chosen depending on the ing algorithm for the fuzzy model is trained by using
domain. Based on Wang's (1992) discussion, fuzzy the EBP-type algorithm. Since the structure is much
modeling with membership functions of scaled Gaus- di€erent from the standard EBP, modi®cation should
sian functions is actually a universal approximator be made. In order to clearly describe the learning al-
that can approximate any non-linear input±output gorithm, the variables used are de®ned in the Nomen-
data arbitrarily well on a compact set. Thus, three par- clature.
ameters of a scaled Gaussian function should be deter- The fuzzy model, in the form of an ANN, can be
mined. For each group or rule, the mean of the group derived from the following equations:
is set to be the center of the membership function. The  e 2
1 x i ÿci,
width of the membership is set to be three times the ÿ2 si, j
j

standard deviation of each group. Membership function: mei, j ˆ si, j e 3†

3.4.1.3. Consequence determination. The consequence, Y

iˆn
which is the control action of each inference rule, is Firing strength: We‰jŠ ˆ mei, j 4†
determined for each group, or rule, by using re- iˆ1, j2‰l, mŠ
gression.
X
we‰jŠ fe‰jŠ
3.4.2. Self-adjusting stage (SAS) ‰jŠ
The basic idea of this stage is to employ the learning Final output: Oe ˆ X 5†
we‰jŠ
capability of the ANN to update the parameters,
‰jŠ
which have already been predetermined by the SOS, of
fuzzy modeling. Consequently, fuzzy modeling should !
be represented as a form of an ANN. The proposed X
Regression model: fe‰jŠ ˆ a‰jŠ, i x ei ‡ b‰jŠ 6†
architecture of the fuzzy model used is based on Taka-
i
gi's (1983) fuzzy system. Then the fuzzy model is
trained by using the EBP-type algorithm. Since the EBP-type algorithm is employed to self-
The proposed fuzzy ANN consists of ®ve layers. adjust the parameters, the inference rules are updated
Layer 1 is the input layer which consists of the real- so as to minimize the cost function E which is de®ned
valued input variables. Every node in layer 2 is the as:
value of the membership function:
ÿ
1X e
ÿ
1 xÿc 2 Eˆ z ÿ O e †2 7†
mA x† ˆ se 2 s 1† 2 e

where x is the input variable and s, c, and s are the where e is the training example number. Each par-
254 R.J. Kuo / Engineering Applications of Arti®cial Intelligence 13 (2000) 249±261

ameter of the control action function is updated by an 4. Experimental setup

amount proportional to the partial derivative of E
with respect to that parameter. The updated learning In order to evaluate the proposed system, the exper-
rule is of the form iments were run using a 20HP LeBlonde lathe. Three
types of sensor: force, vibration, and acoustic emission,
@E
Wi,t‡1†
j ˆ Wi,t†j ÿ Z 8† were employed. For forces in the feed, radial and main
@wi, j cutting directions, a three-axis Kistler Z3392/b piezo-
electric force dynamometer was used, while two PCB
where wij is the updated parameter. All the above-men-
accelerometers were employed for vibrations in the
tioned equations are used to ®ne-tune the parameters
feed and main cutting directions. A physical acoustics
which have been determined by the SAS of the fuzzy
emission sensor was located at the center of the tool
inference system. Thus, this section shows how to
for monitoring acoustic emission signals. Force sen-
employ the EBP-type learning algorithm for fuzzy
sors, vibration sensors, and an acoustic emission sen-
inference system. For a detailed derivation, refer to
sor were connected to the Kistler three-channel model
(Kuo, 1994).
5804 charge ampli®er, PCB charge ampli®ers, and
DECI AE preampli®er, respectively. Thereafter, force
3.5. Tool/work distance compensation
sensory outputs were connected to a National Instru-
ments acquisition board which was connected to an
The ®nal decision, the amount of tool wear which is
IBM compatible PC with the Labview software pack-
made by the multi-sensor integration technique, is used
age, while vibration sensory outputs were connected to
to decide whether the tool is fresh or worn out and
a Tektronix 2630 Fourier Analyzer which was con-
how to adjust the distance between the tool and work-
nected to an IBM compatible PC with a Fourier Ana-
ing material. The approach used to keep the distance
lyzer package. The acoustic emission output was
between the working material and tool constant is to
connected to an ANALOGIC acquisition board which
adjust the depth of cut during the operation of CNC
was connected to an IBM compatible PC with ANA-
turning machines. For example, in the CNC turning
LOGIC FAST Series package. The sampling rates of
machine program, the depth of cut is adjusted accord-
forces, vibrations, and acoustic emission were 3 KHz,
ing to the following equation:
25.6 KHz, and 1 MHz, respectively, after carrying out
D t ‡ 1 † ˆ D t† ‡ DD t† 9† some research (Chryssolouris et al., 1992; Rangwala
and Dornfeld, 1990; Karmarthi et al., 1991) and doing
where some tests. In addition, in order to make sure that all
W(t ) = the amount of tool wear at time t the three acquisition systems were triggered at the
D(t ) = the depth of cut at time t same time, an automatic trigger was connected to the
DW t† ˆ fW t† ÿ W t ÿ 1†g three systems. Once the trigger was initiated, it started
DD t† ˆ DW t† tan clearance angle† all three data acquisition systems simultaneously. This
allowed sensor data to be taken at the end of a cut
for each monitoring iteration. Fig. 2 shows that when and the measured wear correlated to the sensor data
the amount of ¯ank wear is W(t ), then the depth of obtained. Meanwhile, a chip breaker was also mounted
cut should be added by DD t†: If W(t ) is larger than a on the top of the insert in order to avoid the sensors
predetermined constant, then a tool change is made. being damaged by the chips.
The ¯ank wear was measured by a Baush & Lomb
toolmaker's microscope, while surface roughness was
measured using a Federal Systems Pocket Surf. A Star-
rett micrometer caliper was used for measuring the di-
ameter of the workpiece.
In this experiment, all the sensory signals from three
cutting forces in the feed, radial and main cutting
directions, two vibrations in the feed and main cutting
directions, and acoustic emission were collected for
each cut and saved as three ®les. Flank wear of the
tool, diameter of the workpiece, and surface roughness
were then measured o€-line. All of the sensor measure-
ments were sequenced using a common trigger just
prior to the end of the cut as described.
This experiment used SAE 6150 chromium-va-
Fig. 2. Tool/work distance compensation nadium alloy steel as the test workpiece. The work-
 R.J. Kuo / Engineering Applications of Arti®cial Intelligence 13 (2000) 249±261 255

piece's dimensions are 7.5 in. diameter by 36 in. length. 5.1. Data acquisition
The quench and tempered heat treatment procedures
of the workpiece are as follows: The data is obtained from three forces in the feed,
1. heated to 15508F, radial, and main cutting directions, two vibrations in
the feed and main cutting directions, and acoustic
2. oil quenched,
emission. For the purpose of explanation, the cutting
3. tempered at 6008F,
condition and the corresponding tool number are
4. air cooled.
shown in Table 1. In total, there are 15 tools employed
The resultant hardness ranges from 350 to 390 BHN. in this experiment. Each tool has 10±13 samples. The
The Kennamental KSBR-164C tool holder was used total number of samples is 168. However, these 15
for machining, while the cutting insert used was a tools are divided into two groups for training and test-
Kennamental K68 grade carbide insert SPG 422 ing. Thus, the shadowed cells represent those data
mounted on the tool holder. used for training ANNs and multi-sensor integration
The cutting conditions were varied in order to methods. Otherwise, they are for the validation of
obtain more robust data sets. Feed rates were varied multi-sensor integration methods. As a result of this
from 0.0064 to 0.0156 ipr. Levels 0.0064, 0.0088, division, the training samples can cover the range of
0.0112, 0.0136, and 0.0156 ipr were selected. For the cutting conditions. Therefore, the numbers of
speed, three di€erent cutting speeds, 100, 130, and 160 training and testing samples are 89 and 79, respect-
sfpm were used. The depth of cut was kept as a con- ively.
stant, 0.05 in. A full factorial experiment was per-
formed. In total, there were 15 di€erent cutting 5.2. Feature extraction
conditions, or treatments (3 speeds  5 feeds).
The experimental procedures are as follows: Two di€erent analyzers are used to extract the fea-
1. Mount the tool insert and chip breaker on the tool tures from the sensory signal pattern.
holder.
2. Set up the cutting conditions and calibrate the 5.2.1. Time series analyzer
ANALOGIC FAST Series package for the acoustic The sensory signal pattern obtained from the sensor
emission acquisition. is a time series. In the following discussion, the sensory
3. Cut the workpiece for 1 min and initiate the trigger signal patterns mentioned are the magnitudes of the
at the end of the cut, for approximately 55 s, in signals. The test developed by Dickey and Fuller is
order to collect the sensory signals for forces in ®rst applied to test for non-stationarity. For each sen-
three directions vibrations in two directions, and sor, 10 signal patterns are randomly selected. The
acoustic emission. results showed that most of them are stationary. Then,
4. Save the sensory signals for forces, vibrations and by examining the autocorrelation function (ACF) and
acoustic emission in the three di€erent ®les. partial ACF (PACF) of the time series data, AR(5) is
5. Remove the tool insert from the tool holder and able to represent all the sensory signal patterns of
measure the ¯ank wear with the Baush & Lomb force, vibration, and acoustic emission. Since the con-
toolmaker's microscope. stant term is also included in the model, there are six
6. Measure the diameter with the micrometer and sur- total features for all the sensory signal patterns. For a
face roughness using the Federal Pocket Surf. detailed discussion, refer to (Kuo, 1994).
7. Remount the tool insert and chip breaker, and con-
tinue repeating steps 3±6 until severe wear, about 5.2.2. Frequency analyzer
0.018 in., is reached. Since the sensory signal patterns obtained from the
experiments are time domain data, these patterns are
There are 4096 data points for each cutting exper-
iment, but only 1024 data points will be used in the
Table 1
next section, model validation. These sensory signal The cutting conditions and the corresponding tool numbers
patterns are processed for feature extraction.
Feed (in.) Speed (fpm)

100 130 160

5. Model validation results and discussion
0.0064 T1 T6 T11
An on-line estimation system and experimental setup 0.0088 T2 T7 T12
0.0112 T3 T8 T13
have been discussed in the previous two sections. In this
0.0136 T4 T9 T14
section, the data obtained from the physical experimen- 0.0156 T5 T10 T15
tation will be used to evaluate the proposed system.
256 R.J. Kuo / Engineering Applications of Arti®cial Intelligence 13 (2000) 249±261

transformed into frequency domain patterns through Table 2 lists all the computational results and the
the FFT in order to observe the frequency spectrum. mean square error (MSE) values. The radial force with
Then the data is sorted and the ®ve greatest values are a combination feature extraction method of 8.11Eÿ9
selected for each pattern as the features of that pat- MSE value is the least when compared with the other
tern. two analyzers. However, they are very close. Basically,
it is not necessary to choose the combination feature
5.3. Pattern recognition extraction method, since its training time is much
longer than the other two. Therefore, both the testing
In the previous section, the features were extracted and training results indicate that the radial force with
through two di€erent analyzers, time series and fre- frequency features will be the candidate for inte-
quency. The time series analyzer employs AR(5) for all gration. A similar result exists for the feed and main
of the patterns, while ®ve peak values are chosen as cutting forces. However, in this research only one will
the features for the frequency analyzer. However, it is be selected among three di€erent directions for force
very important to ®nd out the better analyzer for this sensors. Thus, the radial force with frequency features
application. Thus, the features obtained from the will be used in the integration. Based on similar con-
above two feature extraction analyzers are used as the cepts, the main cutting vibration using the time series
inputs to the EBPN as well as CPN, and the following analyzers will be selected. For acoustic emission, fea-
three alternatives will be tested for each sensor: tures extracted from the frequency analyzer only, have
the lowest MSE value, 1.71Eÿ8.
1. features extracted by the frequency analyzer;
2. features extracted by the time series analyzer; and 5.3.2. Counterpropagation network
3. features extracted by both frequency and time series The setup of CPN is similar to EBPN. The compu-
analyzers. tational results are shown in Table 3. Thus, the radial
In addition, cutting conditions, feed and speed, and force with time series features will be used in the inte-
cutting time are also employed as features. gration. Based on similar concepts, the main cutting
vibration using the frequency analyzers having the
5.3.1. Error backpropagation network (EBPN) lowest MSE value will be selected. For acoustic emis-
For the features extracted from the time series ana- sion, features extracted from the frequency analyzer
lyzer, the EBPN architecture consists of nine input have the lowest MSE value, 1.05Eÿ5.
units, cutting time, speed, feed, and six time series fea- In the next section, a fuzzy model will integrate only
tures, which are connected to the hidden units which the best sensors' results for force, vibration, and acous-
are connected to one output unit (predicted amount of tic emission. The above results indicate that the EBPN
¯ank wear). Regarding the number of hidden units, can always provide more accurate results than CPN.
some alternatives are testi®ed. The results showed that The only advantage of CPN is that the training time is
18-hidden-unit can provide the smallest MSE value. shorter. Since the computational speed of EBPN is
Therefore, the network structure is 9(input)±18(hid- enough to provide a timely decision, it is selected
den)±1(output). For the frequency analyzer, the EBPN instead of CPN.
architecture consists of eight input units which are
connected to 16 hidden units which are connected to 5.4. Multi-sensor integration
one output unit. Similarly, 16 is the result of imple-
menting some tests. If both of the di€erent analyzers The amounts of tool wear obtained from three sen-
are combined in order to provide the inputs to the
EBPN, the network architecture consists of 14 input
Table 2
units which are connected to 28 hidden units which
The MSE values of EBPN for three feature extraction methods
are connected to one output unit. Based on the above
three di€erent kinds of feature extraction algorithms, Sensors Feature extraction algorithm
the computation is implemented by using the C pro-
gramming language. Before training, the features, cut- Frequency Time domain Combination
ting conditions, cutting time, and measured amounts
Force
of tool wear are normalized within (0, 1). For all the Feed 7.39Eÿ8 2.65Eÿ8 2.11Eÿ8
di€erent cases, the training is considered complete Radial 8.34Eÿ9 9.68Eÿ8 8.11Eÿ9
when the the number of training epochs is 1,500,000. Main cutting 5.89Eÿ7 1.34Eÿ8 9.21Eÿ9
The training rate and momentum are set to 0.1 and Vibration
Feed 8.38Eÿ8 7.33Eÿ8 7.19Eÿ8
0.9, respectively, while the steepness of the activation
Main cutting 5.64Eÿ8 4.74Eÿ8 4.58Eÿ8
function is set to 0.5. A 586 IBM compatible PC is uti- Acoustic emission 1.71Eÿ8 6.82Eÿ7 5.35Eÿ7
lized.
 R.J. Kuo / Engineering Applications of Arti®cial Intelligence 13 (2000) 249±261 257

sors, or ANNs, are combined through three di€erent computational result shows that the value of MSE for
multi-sensor integration methods: multiple regression, the predicted amounts of tool wear is equal to 1.2 
ANN, and FNN, for the purpose of comparison. The 10ÿ7 as the number of epochs is 1,500,000.
predicted and measured amounts of tool wear for
training samples will train these three integration 5.4.3. FNNs
methods. In order to make a comparison of these FNNs consist of two stages, SOS and SAS. Koho-
three integration methods, the value of the mean nen's feature mapping divides the data into three
square error (MSE) is calculated as the criterion. MSE groups in SOS representing small, medium, and large
was also used by Chryssolouris and Domroese (1989) amounts of tool wear, since the wear curve may be
for the purpose of comparison. By using MSE, the ab- easily segmented in this manner for physical reasons.
solute distance between the predicted amount of tool For each group or rule, the mean and standard devi-
wear and the measured, or actual, amount of tool ation are calculated. The center is set to be the mean.
wear can be observed. The width is set to be three times the standard devi-
ation and the scale is set to be one for the initial state.
5.4.1. Multiple regression Moreover, the regression model, which is the conse-
For the regression model, there are three indepen- quence of each rule or group, is also found.
dent variables: predicted amounts of tool wear from Through the above initial set-up, the self-adjusting
radial force, main cutting vibration, and acoustic emis- stage will ®ne-tune all of these parameters, scales, cen-
sion. Meanwhile, the dependent variable is the ters, widths, and coecients of the regression models
measured amount of tool wear. The multiple regression for all of these three rules. Thus, by using the learning
model will also include the constant term. The multiple algorithm proposed in Section 2, these parameters are
regression model is determined by using the MINITAB updated epoch by epoch. After epoch 1000, the MSE
statistical software. The ®tted model is as follows: drops to 2  10ÿ9. If training is continued, MSE will
decrease to 1.32  10ÿ9 when the number of epochs is
y^ ˆ 0:569032x 1 ‡ 0:432042x 2 ‡ 0:004549x 3 10,000. This implies that it is not necessary to train the
FNN for a long time, since FNN is able to learn the
‡ 0:004022 10† training samples in a short time and its accuracy is not
signi®cantly improved if training is continued. The re-
where yà is the predicted amount of tool wear after inte- gression models for each rule before and after ®ne-tun-
gration, x 1 , x 2 , and x 3 are the predicted values from ing are:
the radial force, main cutting vibration, and acoustic
emission, respectively. The MSE of the predicted Rule 1
amounts of tool wear is equal to 4  10ÿ9.
IF
5.4.2. ANNs the amount of tool wear predicted by radial force
For the network which integrates the results of is small AND
radial force, main cutting vibration, and acoustic emis- the amount of tool wear predicted by main cut-
sion, the ANN architecture consists of three input ting vibration is small AND
units and which are connected to six hidden units, the amount of tool wear predicted by acoustic
which are ®nally connected to one output unit. The emission is small
THEN the amount of tool wear is:
Table 3 before y^ ˆ 062963x 1 ‡ 0:22951x 2 ÿ 0:11758x 3 ÿ
The MSE values of CPN for three feature extraction methods 0:000055
after y^ ˆ 0:681134x 1 ‡ 0:219751x 2 ÿ 0:142628x 3 ÿ
Sensors Feature extraction
algorithm 0:00081

Frequency Time Combination

domain Rule 2
Force
IF
Feed 2.44Eÿ5 2.39Eÿ5 9.74Eÿ6
Radial 6.47Eÿ6 6.68Eÿ6 6.18Eÿ6 the amount of tool wear predicted by radial force
Main cutting 7.25Eÿ5 6.83Eÿ5 1.22Eÿ5 is medium AND
Vibration the amount of tool wear predicted by main cut-
Feed 1.64Eÿ5 7.22Eÿ5 1.33Eÿ5 ting vibration is medium AND
Main cutting 8.76Eÿ6 9.55Eÿ6 8.43Eÿ6
the amount of tool wear predicted by acoustic
Acoustic emission 1.05Eÿ5 4.33Eÿ5 1.39Eÿ5
emission is medium
258 R.J. Kuo / Engineering Applications of Arti®cial Intelligence 13 (2000) 249±261

THEN the amount of tool wear is: the amount of tool wear predicted by main cut-
before y^ ˆ 0:268101x 1 ‡ 0:077119x 2 ‡ 0:731269x 3 ting vibration is large AND
‡ 0:000538 the amount of tool wear predicted by acoustic
after y^ ˆ 0:266292x 1 ‡ 0:066931x 2 ‡ 0:681062x 3 ÿ emission is large
0:000221
THEN the amount of tool wear is:
before y^ ˆ 0:0115781 ‡ 0:993291x 2 ‡ 0:004241x 3 ÿ
Rule 3 0:004433
after y^ ˆ 0:0098215x 1 ‡ 1:101938x 2 ‡ 0:0056218x 3
IF ÿ 0:008821
the amount of tool wear predicted by radial force
is large AND So far, three di€erent methods have been developed.

Fig. 3. Measured and predicted amount of tool wear through FNN for cutting SAE6150(370BHN) with Kennamental K68/SPG422 insert.
 R.J. Kuo / Engineering Applications of Arti®cial Intelligence 13 (2000) 249±261 259

In the next section, these three methods will be evalu- samples. It is also identical to the training results,
ated by 79 test samples. since the network trained by radial force features can
converge fastest.
5.4.4. Testing/validation In multi-sensor integration, it is shown that FNN
The testing results via FNN are shown in Fig. 3, can predict the amount of tool wear more accurately
while Fig. 4 shows the pro®les with and without than the other two multi-sensor integration methods
adjusting the depth of cut using FNN results. (multiple regression and ANN) when no sensor is
defective. FNN also outperformed multi-sensor fusion
5.5. Discussions using ANN.
Without doubt, the MSE value of FNN is less than
For time series models, based on the results of non- the multiple regression's MSE value. If the training
stationarity tests, all the sensory signal patterns are samples are divided into several groups and use a mul-
stationary. After examining the ACFs and PACFs tiple regression model for each group, then the result
based on Kendall and Ord (1990), AR(5) is best able will be better than the single multiple regression
to represent most of the patterns for force, vibration, model. Similarly, FNN is very similar to this method
and acoustic emission. Though some of the patterns' which divides the training patterns into groups and
ACFs and PACFs may show that AR(6) or AR(7) is uses a multiple regression model for each group. One
an appropriate model, the majority agree with AR(5). di€erence is that FNN with its membership functions
If the ARMA model is used, it will not improve the can more accurately model the relationship between
speed of learning for an ANN, but will increase the the inputs and outputs. Also, the learning capability of
number of the features. The reason for this is that the FNN can ®ne-tune the parameters in order to improve
coecients from MA are not able to represent the pat- accuracy.
terns, and it can be found that there are no spikes for From the plotting of measured and predicted
ACF. amounts of tool wear, it can be found that some
For pattern recognition, this paper tries to testify amounts of tool wear can not be predicted very well
two di€erent neural networks, EBPN and CPN. The because the hardness of the workpiece is not uniformly
training and testing results indicate that EBPN can distributed. The sudden change in the hardness will
provide much better forecasting results than CPN. signi®cantly change the sensory signal patterns. Mean-
However, the disadvantage of EBPN is that it requires while, a crater is also found under conditions of high
a longer training time. Since the di€erence in MSE speed and/or feed. This will also a€ect the sensory sig-
values between EBPN and CPN is very large, it is not nal patterns. For instance, ¯ank wear will increase the
reasonable to select CPN. In addition, three fuzzy force, while crater wear will decrease the force in some
models developed by Kuo (1995) signi®cantly decrease cases.
the training time. Therefore, in the integration, only Based on the above discussions, FNN, which is a
the results from EBPN are choosen. kind of fuzzy model, is able to predict the amount of
For a single sensor, radial force is more closely re- tool wear more accurately compared with the other
lated to the amount of tool wear based on the testing methods. Though Ko and Cho (1994) have applied the

Fig. 4. Pro®le with and without compensation.

260 R.J. Kuo / Engineering Applications of Arti®cial Intelligence 13 (2000) 249±261

fuzzy model for tool wear monitoring in diamond strategies for integrating sensor information in machining. Annals
turning, the study only considers three wear states (i.e. of the CIRP 38, 425±428.
Chryssolouris, G., Domroese, M., Beaulieu, P., 1991. A statistical
initial, gradual, and severe) instead of the continuous approach to sensor synthesis. Transactions of the North
amount of tool wear. In addition, Chryssolouris et al. American Manufacturing Research Institution of SME, 333±337.
(1992) employ an ANN and statistical method to Chryssolouris, G., Domroese, M., Beaulieu, P., 1992. Sensor syn-
ascertain the continuous amount of tool wear. How- thesis for control of manufacturing processes. ASME Journal of
ever, the amount of experimental data is insucient Engineering for Industry 114, 158±174.
Dan, S., Chattopadhyay, A.B., Murthy, A.S.R., 1996. Force par-
since only seven tool inserts are used. For the current
ameters for on-line tool wear estimation: a neural network
study, there is more sucient experimental data, which approach. Neural Networks 9 (9), 1639±1645.
is enough for both training and testing. Elanayar, S., Shin, Y.C., 1992. Robust tool wear estimation via
radial basis function neural networks. In: Proc. of The Winter
Annual Meeting of the American Society of Machanical
Engineers, 37±47.
6. Conclusions Freeman, J.A., Skapura, D.M., 1991. Neural Networks Ð
Algorithms, Applications, and Programming Techniques.
In this paper, an on-line estimation system applied Addison-Wesley, Reading, MA.
in the area of tool wear monitoring has been proposed. Fukuda, T., Shibata, T., 1992. Hierarchical intelligent control for
The proposed system is able to accurately predict the robotic motion by using fuzzy, arti®cial intelligence, and neural
network. In: Proc. of IJCNN'92, I-269±I-274.
amount of tool wear. The comparison result also indi- Guillot, M., Oua®, A.E., 1991. On-line identi®cation of tool break-
cates that EBPN can provide a better forecast than age in metal cutting processes by use of neural networks. In:
CPN. In addition, utilization of the EBP-type learning Proc. of ANNIE'91, 701±709.
algorithm for a fuzzy model is pretty promising. Hertz, J., Krogh, A., Palmer, R.G., 1992. Introduction to the
The proposed multi-sensor integration method, Theory of Neural Networks and Fuzzy Systems Ð A Dynamical
Systems Approach to Machine Intelligence. Prentice-Hall,
FNN, is found to be more accurate than other
Englewood Cli€s, NJ.
methods including multiple regression and an ANN on Jang, J.-S.R., 1991. Fuzzy modeling using generalized neural net-
the basis of MSE. It signi®cantly reduces the training works and kalman ®lter algorithm. In: Proc. of Ninth National
time when compared with ANN, since the ®rst stage Conference on Arti®cial Intelligence, 762±767.
of FNN, SOS, has divided the training samples into Jang, J.-S.R., 1992. Fuzzy controller design without domain expert.
several groups or rules and the second stage, SAS, In: IEEE International Conference on Fuzzy Systems, 289±296.
Jang, J.-S. R., Sun, C.-T., 1993. Functional equivalence between
only needs to ®ne-tune the parameters. Regarding the radial basic function networks and fuzzy inference systems. IEEE
pro®le of the cutting material, its precision is higher Transactions on Neural Networks 4 (1), 156±159.
since the predicted tool wear value is applied to adjust Kamarthi, S.V., Sankar, G.S., Cohen, P.H., Kumara, S.R.T., 1991.
the depth of cut. On-line tool wear monitoring using a kohonen's feature map. In:
Proc. of ANNIE'91, 639±644.
Kendall, S.M., Ord, K., 1990. Time Series, 3rd ed. Oxford
University Press, New York.
Acknowledgements Khanchustambham, R.G., Zhang, G.M., 1992. A neural network
approach to on-line monitoring of a turning process. In: Proc. of
This research work was ®nancially supported in part IJCNN'92, II-889±II-894.
Ko, T.J., Cho, D.W., 1994. Tool wear monitoring in diamond turn-
by the National Science Council of Taiwan, ROC,
ing by fuzzy pattern recognition. ASME Journal of Engineering
under Contract NSC 85-2213-E-214-006. Their support for Industry 116, 225±232.
has been appreciated. In addition, the experimental Kosko, B., 1992. Neural Networks and Fuzzy Systems Ð A
data were collected in the Machining Research Lab, Dynamical Systems Approach to Machine Intelligence. Prentice-
Department of Industrial and Manufacturing Engin- Hall, Englewood Cli€s, NJ.
eering, The Pennsylvania State University, USA. Thus, Kuo, R.J., 1994. Multi-sensor fusion for intelligent control of manu-
facturing processes using arti®cial neural networks and fuzzy
the support of Professor P.H. Cohen is also appreci- modeling. Ph.D. Dissertation, Department of Industrial and
ated. The author would also like to thank the anon- Management Systems Engineering, The Pennsylvania State
ymous referees for reading the paper and o€ering University, PA.
many helpful comments. Kuo, R.J., 1995. Intelligent diagnosis for turbine blade faults using
arti®cial neural networks and fuzzy logic. Engineering
Application of Arti®cial Intelligence 8, 24±35.
Kuo, R.J., Xue, K.C., 1999a. Intelligent sales forecasting system
References through integration of arti®cial neural network and fuzzy neural
network. Fuzzy Sets and Systems 108 (2), 123±143.
Chryssolouris, G., Domroese, M., 1988. Sensor integration for tool Kuo, R.J., Xue, K.C., 1999b. A decision support system for sales
wear estimation in machining. In: Proc. of the Winter Annual forecasting through fuzzy neural network with asymmetric fuzzy
Meeting of the ASME, Symposium on Sensors and Controls for weights. Decision Support Systems 24 (2), 105±126.
Manufacturing, 115±123. Kuo, R.J., Chen, J.H., Hwang, Y.C., 2000. An intelligent stock
Chryssolouris, G., Domroese, M., 1989. An experimental study of trading decision support system through integration of genetic al-
 R.J. Kuo / Engineering Applications of Arti®cial Intelligence 13 (2000) 249±261 261

gorithm based fuzzy neural network and arti®cial neural network. works. IEEE Transactions on Systems, Man, and Cybernetics 19
Fuzzy Sets and Systems, in press. (2), 299±314.
Lee, C.C., 1990a. Fuzzy logic in control systems: fuzzy logic control- Rangwala, S., Dornfeld, D., 1990. Sensor integration using neural
ler: part I. IEEE Transactions on Systems, Man, and Cybernetics networks for intelligent tool condition monitoring. Journal of
20 (2), 404±418. Engineering for Industry 112, 219±228.
Lee, C.C., 1990b. Fuzzy logic in control systems: fuzzy logic control- Shibata, T., Fukuda, T., Kosuge, K., Arai, F., 1992. Skill based con-
ler: part II. IEEE Transactions on Systems, Man, and trol by using fuzzy neural network for hierarchical intelligent con-
Cybernetics 20 (2), 419±435. trol. In: Proc. of IJCNN'92, II-81±II-86.
Leem, C.S., Dreyfus, S.E., 1992. Learning input feature selection for Takagi, T., Sugeno, M., 1983. Derivation of fuzzy control rules from
sensor fusion in tool wear monitoring. In: Proc. of ANNIE'92, human operator's control actions. In: Proc. of the IFAC
815±820. Symposium on Fuzzy Information, Knowledge Representation
Lin, C.T., Lee, C.S.G., 1991. Neural-network-based fuzzy logic con- and Decision Analysis, 55±60.
trol and decision system. IEEE Transactions on Computer C-40 Takagi, T., Hayashi, I., 1991. NN-driven fuzzy reasoning.
(12), 1320±1336. International Journal of Approximate Reasoning 5, 191±212.
Lin, C.T., Lee, C.S.G., 1994. Reinforcement structure/parameter Tansel, I.N., 1990. Neural network approach for representation and
learning for neural network based fuzzy logic control systems. simulation of 3D-cutting dynamics. Transactions of NAMRI/
IEEE Transactions on Fuzzy Systems 2 (1), 46±63. SME, 193±200.
Lippmann, R.P., 1987. An introduction to computing with neural Tansel, I.N., Laughlin, C.M., 1991. On-line monitoring of tool
nets. IEEE ASSP Magazine 4 (2), 4±22. breakage with unsupervised neural networks. Transactions of
Malakooti, B., Zhou, Y., 1992. An applications of adaptive neural NAMRI/SME, 364±370.
networks for an in-process monitoring and supervising system. Tansel, I.N., 1992. Identi®cation of the pre-failure phase in micro-
In: Proc. of IJCNN'92, II-534±II-539. drilling operations with multiple sensors. In: Proc. of The Winter
Masory, O., 1991. Detection of tool wear using multi-sensor readings Annual Meeting of the American Society of Mechanical
de®ned by arti®cial neural network. In: Proc. of SPIE: Engineers, 23±36.
Applications of Arti®cial Neural Networks II, Vol. 1469, 515± Wang, L.-X., Mendel, J.M., 1992. Back-propagation fuzzy system as
520. nonlinear dynamic system identi®ers. IEEE International
Nakayama, S., Horikawa, S., Furuhashi, T., Uchikawa, Y., 1992. Conference on Fuzzy Systems, 1409±1418.
Knowledge acquisition of strategy and tactics using fuzzy neural
networks. In: Proc. of IJCNN'92, II-751±II-756. Ren-Jieh Kuo earned an MS degree in industrial and manufacturing
Rangwala, S., Dornfeld, D., 1987. Integration of sensors via neural systems engineering from Iowa State University, Ames, IA, USA in
networks for detection of tool wear states. In: Proc. of the Winter 1990 and a Ph.D. degree in industrial and management systems en-
Annual Meeting of the ASME, PED 25, 109±120. gineering from the Pennsylvania State University, University Park,
Rangwala, S., 1988. Machining process characterization and intelli- PA, USA in 1994. Currently, he is an associate professor of Indus-
gent tool condition monitoring using acoustic emission signal trial Engineering Department, National Taipei University of Tech-
analysis. Ph.D. Thesis, University of California, Berkeley. nology, Taipei, Taiwan. His research interests include both
Rangwala, S.S., Dornfeld, D.D., 1989. Learning and optimization of architecture and applications issues in integration/fusion of arti®cial
machining operations using computing abilities of neural net- neural networks, fuzzy logic, and genetic algorithms.