2019 IEEE 3rd Information Technology,Networking,Electronic and Automation Control Conference (ITNEC 2019)

An Intelligent Control System for Grinding

Ding Ning, Duan Jingsong, Liu Chao, Jiang Shuna
College of Mechanical and Vehicle Engineering, Changchun University
Changchun, China
[email protected]

Abstract—Based on expert system ˈ fuzzy logic ˈ neural

network and grinding theory, an intelligent control system for
grinding process was proposed. The system is constructed by
grinding parameter decision support system (DSS), size
prediction control system ˄ SPCS ˅ and roughness prediction
control system˄RPCS˅. The initial grinding parameters are
decided by the grinding parameter decision support system (DSS)
based on neural network. Size prediction control system
˄SPCS˅is constructed by fuzzy control subsystem, deformation
control subsystem and size prediction subsystem. The first and
the second derivative of the actual amount removed from the
workpiece are added into the network input, which can greatly Fig. 1. Schematic diagram of intelligent control system.
improve the size prediction accuracy. The roughness prediction
control system ˄ RPCS ˅ is made from surface roughness
A. Grinding parameter decision support system (DSS)
prediction subsystem and fuzzy neural network controller, which
can adapt grinding parameters in process to decrease the surface First, people enters the workpiece material, original
roughness of machined parts when the roughness is not meeting dimensions, grinding wheel type and other machining
requirements. The experiment of the external cylindrical conditions and the finished dimension, dimensional accuracy,
grinding was implemented. The results showed that this surface roughness and other machining requirements into DDS.
intelligent control system was feasible, and could significantly DDS automatically selects the grinding parameters from expert
increase the grinding quality. library. Besides, DSS can learn by itself during grinding
process, which is shown in Fig.2.
Keywords—intelligent system ; grinding; size prediction control;
roughness prediction control

I. INTRODUCTION
Grinding is much more complicated than turning, milling,
drilling and other machining processes, and is one of the
difficult control processes although extensive research has
already been done [1,2,34,5,6]. It is non-linear, random and
indeterminate 㸪 and is almost impossible to build an exact
mathematical model. So the traditional control system is
difficult to realize in the process. Adjustments on the grinding
process are often on the basis of the operator’s experience and Fig. 2. Construction of grinding process parameter decision support system
skill. (DSS).

In this paper, we proposed an intelligent control system. DSS is an expert system based on neural network[7]. In the
We applied this system to the practical cylindrical traverse traditional expert system, the difficult problems are storing and
grinding process, and experiments yielded satisfying results. Learning the knowledge. Since the neural network has self-
learning and self-organizing ability, and not mentioned
II. CONSTRUCTION OF INTELLIGENT CONTROL SYSTEM excellent abillity of inference involving memory association
The schematic diagram of the proposed intelligent control and image thinking, we build the network parameter decision
system is shown in Fig. 1. It is composed of grinding parameter support system, which can extract learning samples from
decision support system (DSS), size prediction control system experts, predetermined functional formulas and empirical
㸦SPCS㸧and roughness prediction control system㸦RPCS˅. formulas directly, infers the knowledge and learns the
knowledge. DSS can put the optimal grinding process
parameters coming from the practical grinding process into the
knowledge library, and overcome the shortcomings of
traditional expert system.

978-1-5386-6243-4/19/$31.00 ©2019 IEEE 2562

B. Size prediction control system㸦SPCS㸧
HQ
The size is one of the important qualities of workpiece. WG HFQ Fuzzycontrol X
Since the longitudinal dimension is much larger than the radial Grinding
dimension for cylindrical workpiece, the deformation is 9I
extremely prone to be produced during grinding process, which WSQ Deformation control
will decrease the size accuracy. We design a size prediction 9I˄Q˅
control system㸦SPCS˅to control the deformation and the W˄Q˅ process
Size prediction W˄Q˅
size accuracy of cylindrical workpiece before the spark-out X˄Q˅ W˄Q˅
stage. (Q WSQ
WQ
The size prediction control system㸦SPCS˅ is composed
of fuzzy control subsystem, deformation control subsystem and Fig. 3. Size prediction control system˄SPCS˅.
size prediction subsystem, as shown in Fig.3. The deformation
control subsystem uses the optimal adaptive control strategy,
and the size prediction subsystem uses the Elman network C. Roughness Prediction Control System(RPCS)
which has memory. In Fig.3, the sandwheel controlled infeed is The spark-out stage is the final period in grinding process.
u; the table feed velocity is Vf ; the workpiece actual removed In this stage, workpiece size is hardly changed, while the
amount is t(n); the workpiece prediction removed amount workpiece surface roughness is greatly reduced. Roughness
which is the size prediction subsystem output is tp(n); the control is mainly implemented in this spark-out stage. The
grinding system output in the n-1 point is t(n-1); the first and proposed roughness prediction control system is shown in
second derivative of t(n-1) are tp’(n-1) and tp’’(n-1), Fig.4.
respectively; the difference of t(n) and tp(n) is En ; and the
workpiece desired removed amount is td; the error between t(n)
G5 GID
and td is e and the error change rate is ec. The size prediction nw/nsˈ FNN controller Grinding process
control works as follows: N Y tˈVa
ID
First, the grinding parameter decision support system (DSS) R(Rp)>Rd
selects the optimal Vf and u , and enters them to the grinding nw/nsˈ
Roughness roughmeter
process. Second, t (n-1), t’(n-1), t’’(n-1), u(n-1) and Vf(n-1) are aplied 5G tˈVa
prediction
to the size prediction subsystem. Then tp(n) can be obtained, and
Rp
is entered to the fuzzy control subsystem. Next, the error e and % 5
5 $
the error change rate ec can be got by the followings:

Fig. 4. Roughness prediction control system˄RPCS ˅.

e(n)= td(n) –tp(n) (1)
It is composed of fuzzy neural network (FNN) controller,
ec(n)=e(n)-e(n-1 (2) surface roughness prediction subsystem and roughness
e(n) and ec(n) are inputs and the controlled infeed u is the measurer. The surface roughness prediction subsystem is an
output in the fuzzy control subsystem. At last, we apply u to intelligent system with fuzzy neural network. In Fig.4, the
the optimal grinding process. The procedures cycle mentioned workpiece removed amount is t; the grinding vibrate signal
above are repeated until the desired size is obtained. amplitude is Va; table feed is fa; the workpiece revolution is nw;
the grinding wheel revolution is ns; the measured roughness is
Elman network is used in the size prediction subsystem. Its R; the roughness prediction subsystem output is Rp; the desired
context layer memorizes the forepasted state of hidden layer, roughness is Rd; the roughness change is dR.
and is entered to the hidden layer together with the input of
network in the succeeding time. It can serve as the time-delay The intelligent system predicte and controlle the surface
arithmetic operator. Hence, the Elman network has the ability roughness as follows.
of dynamic memory. At the beginning, we switch A on, and switch B off. The
In the size prediction subsystem, we input the first and roughmeter measures the workpiece surface roughness, and
second derivative of t(n-1) into Elman network, then the change contrast the measured roughness and the desired roughness.
tendency of the workpiece removed amount can be given, We put dR, nw, ns, Va and t into the fuzzy neural network(FNN)
which can improve the prediction accuracy. controller if the grinding roughness is not satisfied. Then the
FNN controller output the table feed deviation dfa. Grinding is
The first and second derivative can be described as the processing on the changed table feed (fa+ dfa). Meanwhile, we
following equations in a disperse system. They are the system measure the workpiece surface roughness by the roughmeter
practice output forward difference. and get the roughness R at this time. Next, R , nw,, ns, , fa , Va
t’㸦n-1㸧= t㸦n㸧- t㸦n-1㸧 (3) and t are entered into the roughness prediction subsystem. The
roughness prediction subsystem receives the above data as
t’’㸦n-1㸧= t㸦n㸧-2t㸦n-1㸧+ t㸦n-2㸧 (4) fuzzy neural network training data. With the growth of train
data number, the prediction subsystem accuracy is getting

2563
higher and higher. When the prediction accuracy fulfills the inference. The rule form is as followings: If x1=A1㸪x2=A2㸪
requirement, the switch B is on, and the switch A is off. The x3=A3㸪x4=A4㸪then fj = pjx1 + qjx2 + rjx3 + sj x4 +tj
roughness prediction subsystem begins to work. The predicted
roughness and the desired roughness are contrasted. The FNN FNN Controller. We control the surface roughness by
controller applies a new table feed to the grinding process if the FNN in the spark-out stage. In this stage, the workpiece
roughness is not reached the standards. The procedures cycle removed amount t is very small and it’s difficult to accurate
mentioned above are repeated until the desired roughness is control by drive motor. Meanwhile, widely studies prove that
obtained.>@. the table feed fa has a great effect on the surface roughness.
Roughness Prediction Subsystem. Many experts have Therefore, we control the workpiece surface roughness by the
studied the relationship between grinding parameters and table feed fa in the proposed FNN controller. The FNN is a five
surface roughness. It is known that roughness varies directly layer fuzzy neural network. As shown in Fig.4, the roughness
with the workpiece revolution nw, table feed fa and the deviation dR, the ratio of the workpiece revolution nw and the
workpiece removed amount t and inversely with the grinding grinding wheel revolution ns, the vibrate amplitude Va and the
wheel revolution ns. From experiments, we find that the vibrate workpiece removed amount t are inputs, and the ajusted value
signal amplitude Va increases in proportion to surface of the table feed dfa is the output.
roughness. III. EXPERIMENT
Taking into account the mentioned above relationships, we
proposed the following equation, Tableĉlists some of the experimemtal conditions. The
desired roughness was Ra0.35μm, and diameter was
Ф24.88±0.005mm. The experiment set up is pictured in Fig.6.
Ra=R(nw/ns)x(tp) y (fa)z (Va)m. (5)

Where, R, x, y, z and m are the experimental constant. That TABLE I. EXPERIMEMTAL CONDITION

means that they are established on the case-by-case basis grinding wheel GB70RAP400 h 32 h 203
although they are constant duiring an individual experiment. In
coolant Common oil water emulsion
generally speaking, the values of R, x, y, z and m in one
Workpiece material quenched 45 steel
grinding condition cannot be used in other conditions. Fuzzy
neural network is helpful in solving this kind of problem. We grinding machi ne modified MMB1320B
take logarithms on both sides of Eq.5. The following equation
can be obtained.
Supposing we can get the following
lgRa=R1+xlg(nw/ns)+yLgt+Zlgfa+mlgVa (6)
Where, R1=lgR.
We build the fuzzy neural network of grinding surface
roughness. The FNN inputs are the logarithm of revolution
ratio nw/ns, the logarithm of workpiece removed amount t, the
logarithm of table feed fa and the logarithm of vibration
amplitude Va, and output is the logarithm of the roughness, Fig. 6. Experiment set up.
shown in Fig.5.
The workpiece finished size under intelligent prediction
control is shown in Fig.7.

Fig. 5. Architecture of the fuzzy neural network.

It can been seen from Eq.6 that the roughness logarithm

and the logarithm of nw/ns, t, fa and Va are lineary relations.
Therefore, R1, x, y, z and m can be got by T-S type fuzzy
Fig. 7. Workpiece finished size.

2564
 Eighteen tests of the roughness intelligent prediction IV. CONCLUSIONS
control were conducted. Tableϩlists the grinding parameters An intelligent control system in grinding process was
in 18 tests. The experiment results of roughness intelligent proposed here. It has higher adaptation ability, and is feasible
prediction control are shown in Fig.8. In Fig.8, PFNNC is the in practical grinding process. The intelligemt system can be
final roughness controlled by the proposed roughness popularized and applied to other processing processes, and will
intelligent prediction control system. Fig.8 shows that the play a positive role for the advancement of manufacturing
proposed roughness intelligent prediction control system technology.
performed 16 successes out of 18 tests.
ACKNOWLEDGMENT
TABLE II. GRINDING PARAMETERS IN 18 TESTS This work was supported by the Jilin Province Education
nw fa tp Department Scientific and Technological Research Project
[r/m] [mm/r] [mm] during National Thirth Five-Year Plan (Grant No.
1 125 8 0.015 JJKH20191189KJ).
2 125 8 0.025 REFERENCES
3 125 15.2 0.015 [1] Michael N. Morgan,Rui Cai,Andrea Guidotti,David R. Allanson,J. L.
4 125 15.2 0.025 Moruzzi,W. Brain Rowe ,ĀDesign and implementation of an intelligent
5 125 24.8 0.015 grinding assistant system, ā International Journal of Abrasive
Technology; 1752-2641; vol.1; no.1; p.106-135, 2007.
6 125 24.8 0.025 [2] Zhou, P.,Chai, T.,Sun, J. “Intelligence-Based Supervisory Control for
7 170 8 0.015 Optimal Operation of a DCS-Controlled Grinding System ā , IEEE
8 170 8 0.025 transactions on control systems technology: A publication of the IEEE
Control Systems Society,1063-6536; vol.21; no.1; p.162-175,2013.
9 170 15.2 0.015
[3] Manoj Kumar,Jyoti Raman,Priya Priya , Ā Particle Swarm
10 170 15.2 0.025 Optimization: Intelligent Neural Network System (PSOINNS) Based
11 170 24.8 0.015 Multi-Objective Optimization of Surface Grinding Operations ā ,
International journal of materials forming and machining processes;
12 170 24.8 0.025 2334-4563; vol.2; no.1; p.54-87, 2015
13 200 8 0.015 [4] B. W. Kruszyski; P. Lajmert. An Intelligent Supervision System for
14 200 8 0.025 Cylindrical Traverse Grinding. CIRP Annals 2005.Vol54: pp305-308
15 200 15.2 0.015 [5] Kazuhito Ohashi; Guifu He; Shinya Tsukamoto. Improvement of
Machining Accuracy in Micro Cylindrical Traverse Grinding. Key
16 200 15.2 0.025 Engineering Materials,Vol 329: pp 39-44, 2007
17 200 24.8 0.015 [6] Li GuoFa ˈ Wang LongShan.Simulation and Optimum for Plunge
18 200 24.8 0.025 Grinding. Key Engineering Materials. Vol202-203:pp219-226,2001
[7] William W. Melek; Alireza Sadeghian. A Theoretic Framework For
Intelligent Expert Systems In Medical Encounter. Expert Systems.
Vol.26:pp82-99, 2009
[8] N. Ding: Study of Intelligent Cylindrical Traverse Grinding.(Ph.D., Jilin
University, China 2004),pp.115-117.

Fig. 8. Workpiece finished roughness.

2565