Quality Assurance in Machining Process Using Data Mining

Tzu-Liang (Bill) Tseng* Terry Leeper Chaitanya Banda Stan M. Herren James Ford
Dept. of Architecture and Dept. of AMS Dept. of Computer Sciences Dept. of AMS Dept. of AMS
Manufacturing Sciences Western Kentucky Western Kentucky Western Kentucky Western Kentucky
Western Kentucky University University University University
University Bowling Green Bowling Green Bowling Green Bowling Green
Bowling Green, KY 42101 KY 42101 KY 42101 KY 42101 KY 42101
Abstract
In this paper, the problem of quality assurance and predicting outcomes in machining process applications
is discussed. This study is focusing the use of data mining technique called rough set theory to derive the
decision rules from historical data to be used for identifying factors that contribute to surface roughness
quality. The proposed rule-composing algorithm and rule-validation procedure were validated by historical
data and the results indicated the practical viability of rough set theory approach for predicting the
acceptance level of surface roughness during machining process.

Keywords: Data mining, Rough set theory, CNC machining, Surface roughness, Quality assurance

1. Introduction
Traditionally, the quality of a machined product has been measured, based on the specifications of the
products, after the machining process has been completed. However, this post-process inspection has
several shortcomings: (1) the manufacturing cost has already been incurred when a non-conformance is
detected, (2) it is difficult to isolate the cause(s) of the defect, (3) there is a significant time lag between
discovery of the defect and corrective action, and (4) rework of any feature increases manufacturing cost
and can be very difficult to accomplish. Today, efforts of manufacturers are shifting from the post-process
inspection to improved monitoring of the manufacturing during the process, by utilizing sensors and other
measurement devices to effectively control the process. Improvements in machining precision can only be
accomplished by the development of manufacturing systems that are capable of monitoring active
processes. The monitoring process control reduces scrap, rework, lead-time, and conventional non value-
added inspection activities, thereby increase the system’s productivity. Furthermore, an expert machinist
can detect and/or predict deteriorating cutting conditions through the use of his/her senses so that
appropriate corrective actions can be taken before the part quality is lost. This level of expertise requires
many years of experience to acquire, and does not guarantee that cutting conditions will be monitored most
effectively or consistently, to ensure that decisions about changes in the process can be made in a timely
manner. Consequently, process factors that affect part accuracy and quality during machining processes are
quite complicated, and the effective control of these factors is beyond the capabilities of even the most
experienced operator. Computer numerical control (CNC) of machine tools only minimizes a small portion
of these errors since CNC often does not utilize sensor data to compensate for errors that are generated by
changes in the cutting process. Specifically, it is rather difficult to predict machining output (e.g., surface
roughness).
This study focuses on the use of the data mining technique called rough set theory (RST) [1] for
investigating the factors that contribute to surface roughness quality, and developing intelligent control
system to be used in the turning operation process. The RST is a viable approach for extracting meaningful
knowledge and making predictions for an individual data object rather than a population of objects [2].
Furthermore, none of the previous approaches is able to identify “defectives” and “significant factors”
simultaneously. Most of the statistical tools must make several statistical assumptions such as normality of
distribution and constant variance of the variables, etc. In contract, the RST based approach does not
strictly require that the data is a normal distribution, or linearized. In this paper, the rule-composing
algorithm and the rule-validation procedure were proposed and have been validated with the historical data.
2. Fundamental of Data Mining and Rough Set Theory
Data mining is the process of extracting and refining knowledge from large database [3, 4, 5]. The
extracted information can then be used to predict, classify, model, and summarize the data being mined.
RST is a fundamental theory of data mining. This theory was originated by Pawlak [1] and was developed
to classify imprecise, uncertain, or incomplete information or knowledge expressed in terms of data
acquired from experience, therefore, it complements the fuzzy set theory [6]. The rough set approach is
suitable for processing qualitative information that is difficult to analyze by standard statistical techniques
[7]. It integrates learning-from-example techniques, extracts rules from a data set of interest, and discovers
data regularities [8].
RST has been applied to address a variety of problems [9]. These applications included (1)
representation of uncertain or imprecise knowledge; (2) empirical learning and knowledge acquisition from
experience; (3) knowledge analysis; (4) evaluation of the quality of the available information, with respect
to its consistency, and the presence or absence of repetitive data patterns; (5) identification and evaluation
of data dependencies; and (6) approximate pattern classification.
The RST is introduced as an extension of fuzzy set theory for the study of intelligent systems that are
characterized by incomplete information and applied to classify imprecise, uncertain, or incomplete
information, or knowledge, the knowledge is expressed in terms of data. Indeed, the RST is an effective
tool for multi-attribute classification problems. In RST, data is expressed in a decision table in which each
row represents an object and each column represents an attribute. Formally, the decision table is
represented by an information function [10] as follows:
S =< U , Q, V , f > (1)
where U is a finite set of objects, Q is finite set of attributes,
V = ∪ Vq and V q is domain of the attribute q, and f : U × Q → V is the total decision function such
q∈Q

that f ( x, q ) ∈ Vq for every q ∈ Q , x ∈U .

The main theme of RST is concerned with measuring what may be described as the “ambiguity” that is
inherent in the data. In RST, the essential distinction is made between objects that may definitely be
classified into a certain category, and those that may possibly be classified. Considering all of the decision
classifications yields to what is referred to as the “quality of approximation,” that measures the proportion
of all objects for which definite classification may be achieved. A rough set can be described as a collection
of objects that in general cannot be precisely characterized in terms of their values of their sets of attributes,
but can be characterized in terms of lower or upper approximations [11]. The upper approximation includes
all objects that possibly belong to the category while the lower approximation contains of all objects that
definitely belong to the category.
As each object is characterized with attributes, then discovering the dependencies between attributes,
and detecting the main attributes is of primary importance in RST. Attribute reduction (reduct) is one
unique aspect of the rough set approach. A reduct is a minimal sufficient subset of attributes, which
provides the same quality of discriminating concepts as the original set of attributes. Consider the five
objects in Table 1 each with four input features and an output feature (outcome).
Table 1. Example Data Set
Object No F1 F2 F3 F4 O
1 0 1 0 2 2
2 0 0 1 3 0
3 0 1 1 1 1
4 1 2 2 0 1
5 0 0 0 1 2
O: Not Applicable, 1: Low, 2: Medium, 3: High
To derive the reducts, consider the first feature F1. The set of objects corresponding to the feature value
F1 = 0 is {1, 2, 3, 5}. This set {1, 2, 3, 5} cannot be further classified solely using the relation F1 =0. It is
discernible over the constraint F1 =0, which is expressed as [x][ F1 = 0] = {1, 2, 3, 5}. For the objects in set
{1, 5} the output feature is O = 2, for object 3 the output feature is O = 1 and for object 2 the output feature
is O = 0. Therefore, additional features are needed to differentiate between O = 0, 1, or 2. By applying this
concept, the classification power of each feature can be evaluated. For instance, the feature value F1 = 1 is
specific to O = 1. This discernible relation can be extended to multiple features, e.g., [x][F1 = 0] ∧ [F2 = 1]=
{1, 3} and [x] [F1 = 0]∨ [F2 = 1] ={1, 2, 3, 5}, where ∧ and ∨ refers to “or” and “and”, respectively.
2.1 Reduct Generation
Most of the rough set based approaches may generate more than one reduct for an object. This paper
adapts the reduct generation procedure proposed by Pawlak [12] and presents it in the form of the reduct
generation procedure (Figure 1). The reduct generation procedure enumerates all possible reducts with one,
two and three features that are presented in Table 2.
Step 0: Set object number i = 1
Step 1: Select object i and find a set of reducts with one to m - 1 features
Step 2: Set i = i +1. If all objects have been considered, go to Step 3; otherwise go to Step 1
Step 3: Terminate the algorithm and output the result
Figure 1. Reduct Generation Procedure

Table 2. Partial Reducts for Data in Table 1

Object F1 F2 F3 F4 O Reduct Ui F1 F2 F3 F4 O
No. No.
1 0 1 0 2 2 1 1 x x 0 x 2
2 x x x 2 2
3 2 0 x 0 x 2
4 0 x x 2 2
5 x 1 0 x 2
6 x 1 x 2 2
7 x x 0 2 2
8 3 0 1 0 x 2
9 0 1 x 2 2
10 x 1 0 2 2

3. The Algorithm
Feature sets are used for predicting an object’s outcome with the algorithms. We adapt the reduct
generation procedure proposed by Pawlak [13]. In RST, the essential distinction is made between objects
that may definitely be classified into a certain category, and those that may possibly be classified. The data
set is randomly divided into the training set and the testing set. The rule-composing algorithm is developed
to derive rules from the training set. Then, we also propose the procedure to validate the derived rules on
the basis of the testing set.

3.1 The Rule-Composing Algorithm

The rule-composing algorithm is for the training set while the rule-validation procedure is for the test set.
Step 1. Define a proper feature set and a target file.
Step 2. Examine each object in the set for completeness. If the object is incomplete, then delete the object
from the file; repeat through all objects.
Step 3. Determine the final rules from the candidate decision rules that are generated by the reduct
generation procedure. If all of the candidate decision rules are satisfied then go to Step 6;
otherwise go to Step 4.
Step 4. If the candidate decision rules are satisfied then transfer the rules to Step 5; otherwise, restore the
objects associated with unsatisfied rule and go to Step 3.
Step 5. Collect all of the satisfied decision rules and go to Step 6; otherwise go to Step 4.
Step 6. Stop and output the results.
Note that Step 1 is critical to obtain high-accuracy outcomes generated by the algorithm.

3.2 The Rule-Validation Procedure

Step 1. Compare each decision rule, derived from the rule-composing algorithm, with each new
object from test set. Calculate how many objects are matched with the rule.
Step 2. Repeat comparison of the decision rules with objects from test set until
no decision rule is left
Step 3. Calculate the accuracy of each rule by using the total matched objects (for each rule)
divided by summation of total correctly matched objects and total incorrectly matched
objects. If accuracy of the rule is greater than a predefined threshold value (e.g., 60%) then
go to Step 4; otherwise, remove the rules.
Step 4. Stop and output the results.
4. An Empirical Study
4.1 Problem Structuring and Data Set Description
Data Mining Modeling
Output: Identification of
significant factors

Feature of the machining

process
*work piece *feed rate
*cutting speed *cutting tool
*depth of cut *tool nose radius
*stiffness *cutting force

Machining CNC Control Surface

Turning
Variables Mechanism process Roughness

Feedback
Figure 2. Structure of the closed loop turning operation process

The historical data were analyzed. Specifically, conditions of the factors, which could meet the
required surface roughness, were identified. Those specific factors will be used to develop the intelligent
control system and also, can be used by industry to optimize the surface roughness of machined metal (e.g.,
aluminum, steel) components. Each information object was described with the following eight features: F1
through F8 and outcome O (see Table 3). The work-piece materials include three different types, 6061-
T6aluminum, 7075-T6 aluminum, and 4140 steel. The 3-D drawing of the part geometry is shown in Figure
3. The surface roughness was measured along the machine Z-axis. The machining has been performed on
Cincinnati Hawk CNC Turning Center. The effects of cutting speed, depth of cut, machine set up-modal
stiffness, feed rate, cutting tool, tool nose radius, and resultant cutting force on the performance of surface
roughness estimation were studied. After roughing and semi-finishing operations, the surface roughness
was measured by means of a Taylor Hobson surface profilometer. The contents of the outcome are
recorded in binary format. “One” means surface roughness is acceptable, while “Zero” means
unacceptable. The significant factors, which have impact on the quality of surface roughness, were
determined through the rule-composing algorithm and the rule-validation procedure. The decision produced
by the algorithm and procedure became the decision rules that were stored in the process control system.

Figure 3. 3-D drawing of the test block

 Table 3. Factor (Feature) set of the machining operation process
Factor Unit
F1 Types of work piece materials N.A.
F2 Cutting speed fpm
F3 Depth of cut in
F4 Machine set up-modal stiffness Lb/in
F5 Feed rate ipr
F6 Cutting tool N.A.
F7 Tool nose radius in
F8 Resultant cutting force lb
Outcome Surface roughness (Ra) µ . in

4.2 Computational Results

The “Rough Set Based Decision Support System” software (see Figure 4) is developed by Western
Kentucky University, with C language and the Common Gateway Interface (CGI) is used as a
communication protocol between the server and client ends. It is implemented using Apache 1.3 web server
to result in a prototype system for remote use.

Figure 4. Rough set application software

Two data sets are used to perform this case study. One is the training data set, which is used to derive
the decision rules; the other is the testing data set, which is used to verify the decision rules. According to
Kusiak (2001), the bootstrapping method suggests splitting the data set according to the following ratio,
0.632 for the training set and 0.368 for the test set. In this case study, training data set was collected for 667
parts and testing data set was collected for 333 parts. 41 out of 667 parts in training set were unacceptable
for surface roughness while 19 out of 333 parts in testing set were rejected. The final decision rules were
illustrated in Table 4.
All of the decision rules that were derived by the rule-composing algorithm were expressed as IF-
THEN rules as shown above. Number of support (see the 3rd column) was recorded from the training set
while the accuracy was verified by the rule-validation procedure from the testing set. One can observe that
all rules include feature1 (types of work piece materials). Therefore, feature 1 is the most significant in this
rule induction. Furthermore, F2, F3, F5 and F8 are significant as well, since they are included in the final
decision rules.
 Table 4. Examples of decision rules
Rule no. Rule expression No. of Accuracy
support
1 IF (F1 = Al 6061) AND (F3 = .2) THEN (D = 1) 102 95%
2 IF (F1 = Al 6061) AND (F5 = .036) THEN (D = 1) 91 100%
3 IF (F1 = Al 7075) AND (F8 = 600) THEN (D = 1) 125 71%
4 IF (F1 = Al 7075) AND (F5 = .017) THEN (D = 1) 75 89%
5 IF (F1 = Steel 4140) AND (F2 = 650) AND (F8 = 200) THEN (D = 0) 7 95%
6 IF (F1 Al 6061) AND (F8 = 600) THEN (D = 0) 8 100%

5. Conclusions
Based on historical data, this study employed rough set method that connects with the causal
relationships between the features of the machining process and acceptance of surface roughness.
According to the rule-composing algorithm and the rule-validation procedure, the RST provides a highly
accurate prediction tool for investigating factors that contribute to the quality of the surface roughness. The
features included in the rules, derived by a rough set algorithm form patterns of different shapes and
therefore, offer a different predictive power. Furthermore, the derived rules constitute the basis for
developing a rule-based intelligent control system for surface roughness quality in the turning operation
process. The analytical effectiveness of a rule-based intelligent control system relies on its ability to explain
its reasoning results. The RST provides important information for acceptance of surface roughness in the
turning operation process. The results showed practical viability of the RST approach for quality control
based on historical data. Future research should elaborate on comparison of accuracy with existing
methods.

Acknowledgements
This research has been partially supported by funds from the National Science Foundation (Grant No.
DMI-0116515) and Kentucky Science and Engineering Foundation (Grant No. KSEF-148-502-02-40).

References
[1] Pawlak, Z., 1982, Rough Sets. International Journal of Computer and Information Sciences, 11(5), 341-
356.
[2] Kusiak, A., 2001, Rough set theory: a data mining tool for semiconductor manufacturing. IEEE
Transactions on Electronics Packaging Manufacturing, 24(1), 44-50.
[3] Berry, M. and Linoff, G., 1997, Data Mining Techniques: For Marketing, Sales, and Customer Support,
John Wiley & Sons, New York.
[4] Dhar, V. and Stein, R., 1997, Seven Methods for Transforming Corporate Data into Business
Intelligence, Upper Saddle River, Prentice-Hall, New Jersey.
[5] Cheung, D.W., Ng, V.T., Fu, A.W., and Fu, Y., 1996, "Efficient Mining of Association Rules in
Distributed Databases," IEEE Transactions on Knowledge and Data Engineering, Vol. 8 (6), 911-922.
[6] Dubois, D. and Prade, H., 1990, "Rough Fuzzy Sets and Fuzzy Rough Sets," International Journal of
General Systems, Vol. 17, 191-209.
[7] Heckerman, D., Mannila, H., Pregibon, D., and Uthurusamy, R., 1997, Proceedings of the Third
International Conference on Knowledge Discovery and Data Mining, Menlo Park, AAAI Press, CA.
[8] Komorowski, J. and Zytkow, J., 1997, Principles of Data Mining and Knowledge Discovery, Springer-
Verlag, New York.
[9] Ziarko, W.P., 1994, Rough Sets, Fuzzy Sets and Knowledge Discovery, Springer-Verlag, New York.
[10] Slowinski, R., 1992, Intelligent Decision Support Handbook of Applications and Advances of Rough
Sets Theory, Kluwer Academic Publishers, Dordrecht, The Netherlands.
[11] Wojcik, Z. M., 1993,"Rough Sets for Intelligent Image Filtering," Proceedings of the International
Workshop on Rough Sets and Knowledge Discovery, Banff, Alberta, Canada, 399-410.
[12] Cabena, P., Choi, H.H., Kim, I.S., Otsuka, S., Reinschmidt, J., Saarenvirta, G., “Intelligent Miner for
Data Application Guide”, available at http://www.readbooks.ibm.com, 1999.
[13] Pawlak, Z., 1991, Rough Sets: Theoretical Aspects of Reasoning About Data, Boston: Kluwer
Academic Publishers.