Professional Documents
Culture Documents
Summary
This paper describes a n intelligent data base system which performs a n intelligent task like a skilled operator. The system consists of a grinding data base. a grinding rule base. a learning module and a reasoning module. The learning module extracts relationships between set-up parameters and results from grinding examples stored in the grinding data base using genetic algorithms. These relationships are expressed in the form of a fuzzy production rule and stored in the grinding rule base. The reasoning module provides suitable dressing and grinding parameters using those rules. A computer simulation is performed to confirm the effectiveness of the proposed system.
2. Concept of Approach
A skilled operator in grinding never remembers all experiences of grinding in his memory. He empirically understands causalities between the input and the output quantities and extracts the relationships between set-up parameters and results through a lot of trials and errors of parameter set-up. His ability obtained in such a way is called "skill". Based on his own skill. he can flexibly cope with parameter set-up for a wide range of grinding specifications and requirements. With respect to set-up of dressing and grinding parameters, operator's skill seems to be superior to any other techniques proposed so far. On the basis of above consideration, this study employs a n approach which imitates the decision-making process of a skilled operator.
379
base together with the subsequent grinding results, a n d utilized for further learning and reasoning.
a .r r
VS
1.0
I
M
VL
4.
2
0
4.1
E m
2
9 cn
0.0 - -
Genetic algorithms are search algorithms based on the mechanics of natural selection and natural genetics. They have attracted a special interest as one of machine learning techniques. The general procedure IS depicted graphically in Fig.:! and reviewed brielly as follows: (1) At beginning, the parameters of a problem are coded a s a finite-length string consisting of 0 s and 1's. 12) The genetic algorithms start with a population of strings which are initialized using random numbers. (3) Fitness of individual strings are calculated according to a n objective function which evaluates the suitability of each string in the given environment. (4) Strings with high fitness are duplicated. While strings with low fitness are eliminated because of the limitation of the population size. This operation is called "reproduction". (5) Among the reproduced strings. some pairs are selected and the same part of each string is exchanged. This operation is called "crossover". The other operation "mutation" is the occasional random alteration of the value of a string position. (6) T h u s a successive population is generated and compared to the previous population. If the previous one is better. then the procedure stops. (7) For further improvement. the new population replaces the previous one. and the procedures from (3) to (6) are repeated.
2 u
Value of Variable
K&nj+
Fig. 3 An example of fuzzy sets and rnenbership functions.
. .
.. .. . . .. . .......... .. .
a . a .
R Z
R2:
T
fld0
R:
RZO
Fig.5 Fuzzy reasoning algorithm. the antecedent part of the rules. and then at which indicates the degree of applicability o the rule i is calculated. The f u z y f sets in the consequent part of rules are modified according to a; and composed as illustrated in the figure. Finally, the gravity center of the composed fuzzy set is taken as the output RZO.
4.3
Implementation
1-
The necessary specifications and the steps required for the implementation of genetic algorithms and fuzzy reasoning algorithm in the system are as follows: (1) Definition of Fuzzy Sets Fuzzy production rules a r e expressed as a chain of elements which represent grinding parameters and results. An example of the fuzzy production rule is shown in Fig.4. Beside basic five fuzzy sets shown in Fig.3. an expanded fuzzy set as shown in Fig.6 is introduced. which includes the range of more than one basic fuzzy sets and makes rules more comprehensive. (2) Representation of Fuzzy Sets with Strings For genetic algorithms, fuzzy production rules have to be coded as strings of binary code. The coding method adopted in this study is illustrated in Fig.6. The length of string is equal to the number of fuzzy sets In an element. and each bit corresponds to one of the basic fuzzy sets. According to this coding method. the rule in Fig.4 can be transformed as follows: 10000 00100 01000. (3) Evaluation Method of Fitness Fitness of rules should be evaluated in the environment which reflects the objective of a problem. In this study, the environnient is implied by grinding examples in the grinding data base. Fitness Fi of a rule i is calculated according to the following equation:
F; = maxl (AL'Ci, ' IMIN,, '(1 -L)i ) , +
j=l
n t
(1)
rt
MINt, =
pyk 1
(3)
where,
n r
11
base.
Fuzzv Set
Stnna
Yes
CEnd]
Fig. 2 General genetic algorithms. Flg.6 Fuzzy sets and strings.
380
In order to confirm the effectiveness of the developed system. a computer simulation was carried out. As a problem to be simulated. cylindrical plunge grinding tasks with a grinding wheel dressed with a single point diamond dresser a r e considered. Each task is specified by the following elements: given specification: dw (workpiece diameter). given requirements: R: (surface roughness), (specific grinding energy as a index of ec grinding burn and surface integrity). set-up parameters: /id (dressing feed]. 6' = [ l - s , ) p e q . (4) ad (dressing depth of cut). / nw [workpiece rotational speed). s, = .S', i L , (5) uw (infeed per rotation of workpiece). ,=I The other elements including grinding wheel. workpiece where. material, stock allowance, and sparkout time are assumed : similarity between rules i and j defined by a ratio sq common to all tasks. Applied fuzzy sets and membership of the number of same elements to the total functions of above seven elements are tabulated in Table 1. number of elements. Consequently. fuzzy production rules in this problem are I : number of rules (i.e.. population size]. expressed in the following form: p,q : exponents. d~ = X i . R Z = x2. and rr = X3 1 IF This strategy decreases a survival probability of a rule which { / d = Yi. nd = Yz. nbv = Y3. and U I = Y4 1 . THEN h a s high similarity to others. where. ( 5 ) Crossover Strategy XI. j = ( VS. 5. M.L. VL I. Y A position of crossover is restricted among borders of elements so that a n invalid element could not be generated. of Simulation Moreover crossover as well a s reproduction should keep 5.2 A Procedure the simulation is shown in Fig.9. a n d procedure of a wide variety of rules. For this purpose. probability of described as follows: crossover between a rule i and a rule j is calculated a s follows: (1) The values of d w . fd. ad. nw. and u w in the ranges shown in Table. 1 were generated using random numbers. (2) The values of R Z and ec were calculated by the following equationsl21191: This strategy suppresses crossover between rules which differ from each other, while it promotes crossover between similar rules with high fitness. ( 6 ) Mutation Strategy Mutation is implemented as change of a fuzzy set in a n where. elenlent. Fig.7 illustrates a n example of the mutation. : diameter of grinding wheel, ds (7) Integration of Rules : rotational speed of grinding wheel. ns After rules with high fitness are obtained. similar rules AI : constants, are integrated so that fuzzy reasoning could be done efficiently ei : exponents. with a small number of rules. When two rules differ from each ~i : error terms. other in only one elenlent. and the f u u y sets in that element Error terms E I in these equations were introduced to take have a common region. the two rules a r e combined a n d transformed into one rule as shown in Fig.8(al. Additionally, when all fuzzy sets of a rule include those of another one as Learning Data shown in Fig.8(b]. the latter rule is included into the former one.
a element k derived from comparison between a rule i and a grinding data I . A : constant. (4) Reproduction Strategy In order to provide suitable set-up parameters for a wide range of grinding tasks. there should be a lot of fuzzy production rules which can be cooperatively applied for a n individual task. In other words. the genetic algorithms should generate a lot of rules with high fitness. For this purpose, a population should keep a wide variety of rules even after many times of reproduction. To meet this requirement. a strategy which decreases the number of similar rules is introduced. This strategy performs reproduction according to a modified fitness Fi' instead of a original fitness F I . Fi is calculated by:
pyt
: a membership grade of
5. 5.1
Simulation
Definition of a Problem
(4) d w , R z , e r
Rule i '
I.
-. .. ...
f
Fig.7 Mutation strategy. Fig.9 Procedure of simulation. Rule i Rule j Table.1 Data structure for simulation.
.
Element
Rule k
(a) Combination
Rule i
Rulej
lnfeed Rate Norkpiece Rotational Soeed Surface Roughness (b) Inclusion Fig.8 Integration of rules. Specific Energy
Rule i
31 8
~-0-0-0-0-3-04-0-0-0-c-o
rnax
i i
Q, fn
fn
rnin
10 Generation
15
20
I
2
s t o c h a s t i c features of R : and c c a s well as a n a c t u a l measurement error into account. and given within -C50/0 of the whole ranges using random numbers. Constants A 1 and exponents ei were obtained through multiple regression analysis on the basis of practical grinding data. (3) T h u s 500 grinding examples were generated as learning data and stored in the grinding data base. Then the learning module was activated to generate fuzzy production rules. 14) Subsequently. for verification of the rules obtained, 500 sets of dw. R:. and er were generated using random numbers. IS) The values of dbb. Rr. a n d ec of verification data were provided to the reasoning module a s requirements of grinding tasks. In response to them. the reasoning module presented the predicted values of pi. ad. nir. and aw. (6) Substituting the values of d w , f d . ad. n w and aw into the equations (7) and 181. the values of Rz and ec were calculated as simulated values(R2' and ec'). (7)RI' and cc' were compared to the required values. and the verification was performed
surface roughness. Fig.10 shows the progress of learning. In the figure. the horizontal axis denotes a generation [i.e.. times of population change), while the vertical axis represents the fitness of rules in a population. It can been seen from the figure that the mean fitness gradually increases with generation and reaches to 0.3 in the 17th generation. With respect to the maximum fitness which corresponds to the best rule. it is obtained in the early 4 t h generation. This result implies t h a t genetic algorithms in the learning module improve the fitness of rules generation by generation. and consequently find appropriate rules efficiently. The 36 rules which were obtained in the 17th generation and had fitness larger than the mean value are shown in Table 2. It can be found from the table that a lot of combinations of seven elements were extracted. Fig. 1 1 shows the comparison between required a n d predicted surface roughness. A fairly good agreement is seen over the entire range. Thus. the result of the simulation verifies t h a t the intelligent d a t a base system is capable of learning adequate relationships between grinding parameters and results on the basis of grinding examples. and providing suitable dressing and grinding parameters for a wide range of grinding tasks.
5.3 Results
6. Conclusions
In this study the intelligent data base system for grinding operations was developed. which consists of the grinding data base, the grinding rule base. the learning module, a n d the reasoning nlodule. For the learning module and the reasoning module. genetic algorithms a n d fuzzy reasoning algorithm were applied, respectively. The system acquires relationships between grinding parameters and results in the form of fuzzy production rule on the basis of grinding examples. Moreover the system utilizes acquired rules, and provides parameters for further grinding t a s k s . The effectiveness of the developed system w a s confirmed through a computer simulation.
References
[ I I Venk. S.: I YY I , A sysrerri Approcich fur Decision Muking rti Producrion Grinding, Annals of the CIRP, Vol.40/1.415.449. (21 Tonshoff, H.K.; Peters. J., Inasaki. I.;Poui. T.: I Y Y ? . Modeiing and Simulation ofCrindirig Processes, Annals of the CIRP, Vo1.41/2. [3j Peters, I.: 1984. Conrrihuiion of CIRP Research io Inclrrsrriui Prohleni in Grinding, Annals of [he CIRP, Vo1.33E. 451468. [4] Sak;ikur;i, M.; Inasaki, 1.: 1992, A Neitrul Nelwork Approach 10 [he DecisionMaking Process for Grinding Operafrom. Aiinals of the CIRP, VoI.JI/I, 353.356 IS] Goldberg. D.E.: 19119, Generic Algorirhms in Search, Opri,nizar~on,und
Machine Learning, Addison-Wesley , 161 Lee, C.C.: 1990, Fuzzy L o g i ~i'n Confro/ S.wrem.! Fuzz); Logic Conlroiler f a r [ / & I / , lEEE Transacuon. VoI.SMC.?O, No. 2, JW-135. [7] Thrift. P ' 1991, Fuzz); Logic Synrhc.ris w i r h Generic Algorirhms, Proceedings of the Fourrh Iiirernarioiisl Conference on Generic Algorirhms.
Morgan Kaufrnann Publisher>. 5(N-513. (81Spechi, D.; Weiss, S 1992. hfotfef-f(asedLearning for Dia,ynostic Tasks, Annals of the CIRP. Vol.4 1 / 1 , 557-560.
'
[9] Mslkin, S.: 1978, Birrninq LUnil fur Surface and Cylindricol Grinding of Sreels, Annals of the CIRP. Vn1.?7/1. 233-236
382