2010 International Conference on Digital Manufacturing & Automation

Adaptively Pattern Recognition in Statistical Process Control Using Fuzzy ART

Neural Network

Min Wangˈ Tao Zan

Key Lab of Advanced Manufacturing Technology, Beijing University of Technology, Beijing, 100124, China
e-mail: [email protected]

Abstract—This paper presents a statistical Process Control

(SPC) method based on Fuzzy ART (adaptive Resonance
Theory) neural network. The Fuzzy ART neural network is
applied to recognize the special disturbance of the

Frequency
manufacturing processes based on the classification on the
histograms. It is shown that the Fuzzy ART neural network
can adaptively learn the features of the histograms of the
quality parameters in manufacturing processes. As a result, the
special disturbance can be automatically detected when a
feature of the special disturbance starts to appear in the
histograms.

Keywords- statistical process control; fuzzy ART; histogram;

pattern recognition
Response variable

I. INTRODUCTION Fig.1 Histogram

Grinding is one of the most common metal machining The histogram graphically shows the center (i.e., the
processes in precision manufacturing. In general, grinding location) of the data, the spread (i.e., the scale) of the data,
processes is the most important machining process, which the skewness of the data, the presence of outliers, and the
must be inspected using Statistical process control (SPC) presence of multiple modes in the data. Generally, The shape
techniques in factories. Based on statistical techniques and of histograms can be categorized into seven patterns:
graphical displays such as control charts and histograms, (1)normal, (2)symmetric, non-normal and short-tailed,
SPC can be used to monitor and ensure the quality of (3)symmetric, non-normal and long-tailed, (4)symmetric and
workpieces and the stability of manufacturing processes. The bimodal, (5)skewed right, (6)skewed left, (7)symmetric with
most important parts of SPC implementation are control outliner. By observing the shape of histograms, it can be
charts and histograms. Their power lie in the ability to detected that whether the manufacturing process is in control
distinguish special disturbance (out of control) with inherent or out of control. The classified patterns on histograms can
variability(in control) based on data analysis and charting. In help users to identify potential variations in manufacturing
this paper, a neural network is presented for recognizing the processes and provide the clue for trouble-shooting. With the
abnormity in histograms and improving the automation of automation development of the manufacturing system, the
SPC implementation in grinding processes. automated recognition of histograms has become the most
The purpose of a histogram is to graphically summarize important task and unavoidable.
the distribution of a univariate data set of quality parameter. In recent years, there is widespread interest in the use of
The most common form of the histogram is obtained by neural networks in manufacturing [1-4]. Successful cases of
splitting the range of the data into equal-sized bins (called applying neural networks for detection the special
classes). The classes can either be defined arbitrarily by the disturbance in control charts and histograms has been
user or via some systematic rule. Then for each bin, the presented by many researchers. General neural networks,
number of points from the data set that fall into each bin are such as back-propagation neural network, utilize a
counted, as shown in fig.1. At the same time, the data should supervised learning scheme, which needs a lot of desired
be normalized. The normalized count is the count in a class histogram pattern sampled from manufacturing processes.
divided by the total number of observations. The relative Because of the complexity and diversity of manufacturing
counts are normalized to sum to one (or 100 if a percentage processes, however, it is very difficult to experimentally
scale is used). In the fig.1, the vertical axis represents the obtain the required training histogram patterns. To overcome
frequency (i.e., counts for each bin), and horizontal axis this problem the use of a Fuzzy ART neural network [5-6]
represents the response variable. This is the intuitive case for special disturbance detection in histograms is proposed in
where the height of the histogram bar represents the this paper. The Fuzzy ART neural network uses an
proportion of the data in each class. unsupervised learning scheme called the competitive

978-0-7695-4286-7/10 $26.00 © 2010 IEEE 160

DOI 10.1109/ICDMA.2010.263
learning algorithm. So, it is not necessary to obtain the Based on the above Monte Carlo method, the data set of
required histogram patterns for training this neural network. the dimension of the ground holes can be simulated. For
The special disturbance can be automatically detected when every pattern, ten sets of the data are sampled with the
unusual histogram pattern appearing. length of 125. At the same time, the data are normalized
In this paper, Monte Carlo method [7] is employed to using the equation:
obtain the desired data. The normalized counts for each bin y (T ) (8)
Y (T )
of histograms are inputted into Fuzzy ART neural network. 20
It is shown that Fuzzy ART network can correctly classify According to the principle of statistics, the distribution of
the patterns of histograms and detect the special disturbance Y(T)should be as follow:
of the manufacturing processes. Once a special disturbance 0.01 2 (9)
appearing during grinding processes, Fuzzy ART network Y (T ) ~ N (0.5, ( ) )
60
can yield a new category, which represents the disturbance The aim of the normalization is to guarantee that the
and is different to other category representing the inherent range of the normalized data is from 0 to 1 and the center of
variability. the data is 0.5. According to “3ı” principle, the range of data
set of the processes in control should be from 0.4995 to
0.5005. Considering the abnormal situation with the data out
II. CONSTRUCTION OF THE DATA SETS USING MONTE
of the limit, the range of the normalized data is from 0.4994
CARLO METHOD to 0.5006. For plotting the histograms, the range of the data
The desired data is difficult to obtain, often requiring is split into 14 equal-sized bins. The count in a bin is divided
Monte Carlo method. Monte Carlo method is a computer by the total number of the set, 125. The sum of the
experimental method that uses random numbers in order to normalized counts is one.
estimate distributions of simulator outputs.
Consider an internal grinding manufacturing problem.
Batches of parts need to be ground. The diameter of the III. THE ARCHITECTURE OF THE FUZZY-ART NEURAL
ground hole is 10mm with a limit of ±0.012mm. For the NETWORK
grinding processes in control, the data of the dimension of Fuzzy ART is a neural network architecture that performs
ground holes should be approximately normally distributed. incremental unsupervised learning of categories recognition
For satisfying with the “3 ɍ ” principle, the distribution is and multidimensional maps in response to input vectors
described as follow: presented in arbitrary order. Fuzzy ART learns to classify
y (T ) ~ N (10, (0.01/3) 2 ) (1) inputs by a fuzzy set of features or a pattern of fuzzy
where, y(T) represents the quality parameter at time T. membership values between zero and one that indicate the
N(10,(0.01/3)2) represents the normal distribution with the extent to which each feature is present. Fuzzy ART
center of 10 and the standard deviation of 0.01/3. At the dynamics are described in terms of analog fuzzy set-
same time, other patterns can be described using Gaussian theoretic operations. The Fuzzy ART architecture is shown
mixtures distribution as follow: in Fig.2. It consists of two subsystems, the attentional
For the pattern of symmetric, non-normal and short-tailed: subsystem, and the orienting subsystem. While in the
1
y (T ) ~ N (10, (0.01/5) 2 )  (2) attentional subsystem familiar patterns are processed, the
3
orienting subsystem resets the neural activity whenever an
1 1
N (10.003, (0.01/5) 2 )  N (9.997, (0.01/5) 2 ) unfamiliar pattern is present as input. The attentional
3 3
For the pattern of symmetric, non-normal and long-tailed: subsystem includes a field, F 0 , of nodes that represents a
2 1 (3) current input vector, a field, F 1 , that receives both bottom-
y (T ) ~ N (10, (0.01/3) 2 )  N (10, (0.01/10) 2 )
3 3 up input from F 0 and top-down input from a field F 2 , that is
For the pattern of symmetric and bimodal: called the category representation field because these
y (T ) ~
1 1
N (10.003, (0.01/5) 2 )  N (9.997, (0.01/5) 2 )
(4) categories represent the clusters to which the input patterns
2 2 belong. At the field F 0 , a complement coding preprocessor
For the pattern of skewed right: transforms the input vector a into the vector I = (a, ac), ac=1-
1 1 a. Associated with each F 2 category node J is a vector w j of
y (T ) ~ N (10.003, (0.01/10) 2 )  N (10.003, (0.01/5) 2 ) (5)
3 3 adaptive weights.
1
 N (9.997, (0.01/3) 2 ) After an input is presented, an index, J, of categories will
3
be selected during the cluster search based on the choice
For the pattern of skewed left:
1 1
function. The orienting subsystem is responsible for
y (T ) ~ N (9.997, (0.01/10) 2 )  N (10, (0.01/5) 2 ) (6) generating a reset signal to F 2 field when the match function
3 3
1
of the chosen category in the attentional subsystem meets
 N (10.003, (0.01/3) )2
the vigilance criterion, ²ę>@ If the reset signal is sent,
3
For the pattern of symmetric with outliner: a new index J of categories will be then chosen. If the reset
5 1 signal is not sent, the formerly selected index of categories
y (T ) ~ N (10, (0.01/3) 2 )  N (10.01, (0.01/10) 2 ) (7)
6 6
represents the best match to current input. The vigilance

161
 A) normal pattern
criterion decides the tolerance of the neural network to the
pattern mismatches. Slight mismatches will be tolerated if
the value of the vigilance criterion is high. Besides the
vigilance criterion, Fuzzy ART dynamics are determined by

Frequency
a choice parameter Į >0 and a learning rate parameter ȕ ę
>@

Normalized data of dimension

B) sysmmetric and short-tailed

Frequency
Fig.2. Architecture of Fuzzy ART neural network
Normalized data of dimension

IV. CLASSIFICATION OF THE FUZZY-ART ON C) skewed right

HISTOGRAMS
In this paper, several different patterns of histograms are
prepared by using Monte Carlo method. For each pattern of
Frequency

histogram, a data set is produced using Monte Carlo method,

which includes a number of 125-dimensional vectors. The
number of points from every 125-dimensional vector that
fall into each bin is counted. And, the normalized counts are
inputted into Fuzzy ART neural network. The input vectors
are transformed from 14-dimensional vector to 28- Normalized data of dimension
dimensional vector by the complement coding preprocessor
D) sysmmetric and bimodal
at the field of F 0. There is a 28-dimensional vector of
weight associated with each category at the field of F 2 . The
vector of weight is divided into two parts, which is
Frequency

respectively connected with the original input vector of the

normalized counts and the complement coding of the
original input vector.
By reasonably setting the parameter of the vigilance
criterion(²), Fuzzy ART neural network has the ability that
all different patterns of histograms can be exactly classified
into different categories. The results of classification of Normalized data of dimension
Fuzzy ART on the pattern of histograms have been shown
E) symmetric with outliner
in fig.3.
As shown in fig.3, there are a number of dots plotted in
each area represents a bin of histograms. Every dot
represents the normalized count in each bin. In every plot
Frequency

for a certain histogram pattern, there are two lines which are
plotted according to the weight vector associated with the
category represents this pattern. By connecting the elements
of the first part of the weight vector using straight line, the
up-line is plotted. The points on the bottom-line are the
complements of the second part of the weight vector. The
Normalized data of dimension
area between up-line and bottom-line represents the
Fig.3 Results of the classification of Fuzzy ART on the pattern of
classification space of the histogram pattern. histograms

162
 The Fuzzy ART neural network is developed using Java ACKNOWLEDGMENT
language for the sake of the convenience of the developed The authors would like to acknowledge the Beijing
application running on the Internet. By utilizing the network Natural Science Foundation and Beijing Science and
technology, the application can on-line monitor and analyze Technology Committee who support this project, and also
the data set of the quality parameters sent through the acknowledge the support of the committee of Beijing
network and rapidly detect the special disturbance of the Education.
machining processes. Fuzzy ART neural network uses an
unsupervised learning scheme and automatically classifies
the data set of the quality parameter into different category.
REFERENCES
When a special disturbance occurring during the machining
processes, Fuzzy ART neural network can automatically
yield a new category differing from the category that [1] A.S. Anagun, “A neural network applied to pattern recognition in
statistical process control,” Computers & Industrial Engineering,
represents the normal pattern of the histograms. Once a new 1998, Vol.35, No.1-2, pp.185~188
category is yielded, an alarm for the abnormality is [2] C. S. Cheng, “A multi-layer neural network model for detecting
generated. changes in the process mean,”. Computers Industrial Engineering,
1995, Vol.28, No.1, pp.51~61
V. CONCLUSION [3] R. Guh, J. Tannock, C. O’Brien, “IntelliSPC: a hybrid intelligent tool
for on-line economical statistical process control,” Expert Systems
In this paper, a neural network, Fuzzy ART, is to be with Applications, 1999, Vol.17, pp.195~212
applied to pattern recognition in the statistical quality [4] B. Kang, S. Park, “Integrated machine learning approaches for
control. Fuzzy ART neural network has the ability to complementing statistical process control procedures,” Decision
automatically detect the special disturbance in Support Systems, 2000, Vol29, pp.59~72
manufacturing processes based on the adaptively [5] M. Pacella, Q. Semeraro, “Alfredo Anglani, Manufacturing quality
classification on the histograms of the quality parameters. control by means of a Fuzzy ART network trained on natural process
data,” Engineering Applications of Artificial Intelligence, 2004,
At the same time, the classification of Fuzzy ART neural Vol.17, pp.83~96
network on the histograms does not need that the previous [6] M. Wang, R.Y. Fei, “On-line Chatter Detection and Control in Boring
knowledge on the distribution of the data set with the Based on an Electrorheological Fluid,”. Mechtronics, 2001, Vol.11,
special disturbance is available. It is very adapted to the No.7, pp.779~792
situation that the disturbance pattern in real manufacturing [7] S. Jeffrey, U. Frederic, “Measuring the stability of histogram
appearance when the anchor position is changed. Computational
processes cannot be manifested by the appearance of Statistics & Data Analysis, 1997, Vol.23, pp335~353
previous patterns.

163