See discussions, stats, and author profiles for this publication at: [Link]

net/publication/221710103

A Robot System for High Quality Belt Grinding and Polishing Processes

Chapter · July 2005

DOI: 10.5772/4680 · Source: InTech

CITATIONS READS

10 945

2 authors:

Bernd Kuhlenkötter Xiang Zhang

Ruhr-Universität Bochum Fourth Military Medical University
336 PUBLICATIONS   2,528 CITATIONS    240 PUBLICATIONS   7,752 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Computational systems biology View project

Arbeit+Innovation View project

All content following this page was uploaded by Xiang Zhang on 21 May 2014.

The user has requested enhancement of the downloaded file.

 A Robot System for High Quality
Belt Grinding and Polishing Processes
Bernd Kuhlenkoetter & Xiang Zhang

1. Introduction
Grinding and polishing are standard operations in the material processing. They are
automated with the help of industrial robots in order to relieve human from laborious
tasks and unpleasant environments and elevate the profitability of production nowadays,
specially in the sanitary fitting industry. However, the systems known at present are
adapted quite costly to other part geometries and operation cycles and are therefore
economically applicable only for large batch sizes. Frequent changing of the robot
program and system parameters also increase the cycle time and the cost of the whole
manufacturing process. This problem will be more outstanding when the operating
surface is a free-formed surface with very complicated geometrical shapes.
This paper describes a robot system which is specially developed for grinding and
polishing free-formed surfaces with high quality and high efficiency due to the linkage of
innovative robot technology, simulation technology and artificial intelligence methods.
The robot system combines some different levels of automation, manual operation, partial
automation and full automation. Highly labor-concentrated jobs will be done full
automatically, for example detecting errors on the workpiece surface within the
manufacturing process. Some modules are partial automated to help the operator plan the
schedule and design the program. For operations that rarely occur a manual interference
will be sufficient. A reasonable remaining of the manual work can keep the low costs of
the robot system which is also acceptable for the enterprises. The robot system is flexible
and suitable for the manufacturing of small and medium batch size, which is a difficult
Open Access Database [Link]

task for current robot-controlled grinding and polishing systems.

The automation in this robot system focuses on two parts. One is the quality assurance
and the other is the process model that suggests the choice of the technical parameters for
a high quality processing. To obtain a high quality surface, the robot system should have
more accurate motion programs in the process optimizing. The new software system is
developed exactly for this purpose. The software helps the operator design the grinding
and polishing paths precisely and efficiently with CAD data of the grinding/polishing
surface. In addition, the simulation of the grinding process will assist the operator to get
more reasonable manufacturing paths and adopt the optimized parameters to the
manufacturing process in order to obtain a high quality surface. The process model is
specially for the grinding process that studies the relation between the various grinding
parameters and the final amount of materials removal from the workpiece surface. The
result of the process model makes the simulation of the grinding process possible. The

Source: Cutting Edge Robotics, ISBN 3-86611-038-3, pp. 784, ARS/plV, Germany, July 2005 Edited by: Kordic, V.; Lazinica, A. & Merdan, M.

755
operator can analyze qualitatively what the surface is like before the workpiece is really
ground. This can direct the operator to modify the grinding path in advance and therefore
reduce the rejection rate of the production processes.
In the next section, the grinding and polishing using industrial robots are briefly
introduced. The system used and the key technologies in the system will be described in
detail in section 3. In section 4 we give an account of our work to simulate the belt
grinding process which is a very important part in the whole system. The sub-system of
defect detection and classification is introduced in the section 5. We end with a summary
and some main future plans of the system.

2. Grinding and Polishing in the Manufacturing Chain

It is the aim of grinding and polishing processing to improve the forming and dimensional
accuracy as well as the surface finishing. Both processes play an important role as they are
at the end of the product chain where processing errors lead to high rejection rates of the
parts.
Specially in the sanitary fitting industry grinding and polishing are widely used to
manufacture a free-form surface. Semi-finished workpieces are normally manufactured by
casting. Through the subsequent grinding and polishing a high-quality shining surface is
produced. In this case, the dimensional accuracy of the workpiece is only playing a
secondary role. Fig. 1. illustrates five phases of producing a water tap fitting from the
rough workpiece to the final product.

Figure 1. Producing steps in the manufacture of fittings. (1) casting, (2) grinding, (3) polishing, (4)
electroplating, (5) end product

Traditionally grinding and polishing processes of such free-form surfaces are done by
manpower. A worker holds the workpiece, feeds it towards the grinding/polishing
machine and moves the workpiece along the paths that are decided subconsciously by
experience, see left side of Fig. 2. In this manufacturing process, the worker has to suffer
from the unpleasant environment, e.g. dirty air and loud noise. Additionally the job is
hard and monotonous work so that the people cannot concentrate on it for very long time,
so they cannot guarantee constant quality of the product. Nowadays, industrial robots are
introduced into the manufacture of free-form surface grinding and polishing to minimize
the costs while optimizing the quality. The robot, replacing manpower, holds the
workpiece and moves along the paths that are predefined by technicians, see right side of
the Fig. 2. To implement this process using industrial robots, the most important is to plan
the grinding/polishing paths and generate the robot programs. Another practical problem
to be handled is the occurring of errors in the previous casting phase. The casting process
is characterized by high resulting dimensional and geometrical tolerances as well as
quality fluctuations like blowholes and pores. These strongly variable starting conditions
of grinding lead to unprofitable rejection rates and a very high manual testing procedures
in the automated grinding and polishing processing. What is even more difficult for the

756
realization of an automated solution is the fact that errors are only detectable after a part
of the fine processing has been done and that the sensitive and strongly shining surfaces
are difficult to measure.

Figure 2. Manual grinding process and robot aided grinding process

The robot aided automation solutions known at present in the fields of grinding and
polishing are especially and successfully used in the sanitary fitting industry. The
profitability of these systems has been granted in the past years despite high national
wage costs. With the advancing globalization of the markets, it is now faced with a high
competition from low-wage countries where the grinding and polishing is done manually.
Besides the grinding and polishing processes, the subsequent steps of manufacturing are
also threatened to be shifted abroad. The cost of the total manufacture process should be
reduced to face global competition. However, the high time and cost requirements for
programming and optimizing have a negative effect on the profitability of industrial robot
aided grinding and polishing cells (Schueppstuhl, 2003). Compared to conventional robot
tasks these high requirements result from the clearly more complex, comprehensive and
more accurate motion programs.
These requirements have of course an even more negative influence if new programming
or adaptations become frequently necessary (Cabaravdic, 2003). The main reasons for this
are an unfavourable ratio of the batch size to the variety of modifications and the
occurring fluctuations in the process due to workpiece tolerances, and other errors in the
upstream manufacturing process. The required level of automation of the procedures
depends on the occurring frequency of the necessary optimizing work. For rarely
occurring process malfunctions a manual interference is sufficient, while for more frequent
occurrences a full automation must be realized. The techniques to be developed for a
manual use differ considerably from a fully automated one. In manual procedures the
focus lies on the efficient interaction with the operator whereas a full automation needs the
development and integration of a complex measuring method, a data processing and
process control.

3. Innovative Robot System

One aspect of the project aims at further developing the process specific offline-
programming system, as approved in practice, in order to reach a higher efficiency than

757
manual programming and optimizing procedure. The nowadays systems are designed for
a universal use and similar to complex 3D CAD-systems what regards their layout and
operation. Processes that do not need an extra path or parameter optimizing, such as
palletizing, assembling or varnishing can be programmed efficiently with such software
systems by high-qualified engineers and technicians in the planning department.
However, there are no appropriate tools available for the grinding and polishing to
optimize the phase directly at the robot cell.
The use of a conventional offline-programming system in grinding or polishing usually
fails because of its complexity in connection with an unsuited qualification of the
operators as well as lack of process specific functions. Therefore a high demand for
according developments arose in this context. The intended system should adapted itself
to small batch sizes and the case in which lots of modifications must often be carried out.
The operator of the robot is in the centre of the decision-taking and should be given PC-
based deciding guidance for the next step and according tools for an efficient program
optimizing.

Automatic adaption of process

Process parameters and generation of Process model
Adoption partial processing programs for the
branching off
Suggestions for the choice of new processing
Process technical parameters for a high parameters
quality processing

Automatic indentification of
defects and evaluation Detection and
Quality classification
Assurance Manual control with structured of defects
defect classification

Figure 3. Central aspects of the system

Another aspect lies in the consideration of disturbing influences that frequently occur and
which must be detected and compensated automatically. In practice, the most common
disturbing influences are the existing defects on the semi-product from the casting phase
or caused directly by the grinding processes.
Therefore, the error detection and classification as well as the deduction of parameter
optimizing strategies through measuring methods and process control must be realized
fully automatically, see Fig. 3.
A software system has been developed for a workshop programming of robot systems that
provides an intuitively operable graphic 3D-user interface. The software provides process
specific optimizing tools, specially for grinding and polishing. The software accepts the
CAD model of the workpiece or its surface as input and helps the operator to generate the
grinding or polishing path conveniently and practically. The software can generate the
according robot programs for users.
With this software, the time required to design the paths is greatly reduced compared to
the universal off-line programming system. It is very suitable for small or medium batch
size manufacture and flexible to the requirements of frequent modifications of workpieces
or paths. Fig. 4. shows the simple generating of grinding paths on the surface of the
workpiece. Additionally, the software is extended by an adaptive consulting centre for the

758
allocation of errors, causes of errors and compensation strategies and an internet
connected process-know-how-database.
This software fills the gap between the multi-functional, however complex, offline-
programming systems used in the planning department and the inefficient possibilities of
the robot control used by the operator for the program optimizing.

Figure 4. User-orientated offline-programming and simulating system (Carat Robotic Innovation, 2004)

A higher tolerance towards changing starting qualities of the workpiece is achieved by

combining image processing/measuring systems, grinding and polishing process models,
and adaptive control technique and intelligent software components. The “see and
evaluate” of processing errors on high shining surfaces is even difficult for the untrained
human eye.
Furthermore, errors in the workpiece material in the process chain of rough grinding,
finish grinding and polishing can be often detected only after a part of the processing has
been done, or at the end.
Therefore, another requirement has arisen for developing a fully automatic working
process chain for the industrial robot-aided grinding and polishing based on the
measurements of an image processing system so that an optimum surface quality is
obtained despite fluctuating starting conditions. Fig. 5. below illustrates the working flow
of the automatic detection and classification system. The central hardware is a closed
vision system that can provide suitable illumination and take photos of the workpiece
surface. The software of the system is responsible for finding the defect position and
classifying the defects by analysing the grayscale photos taken by the vision system.

759
 defect data robot path and
trajectory generation for technical parameters
defect elimination

polished n.i.O.
fitting

defect detection grinding

primary i.O. polishing
machining

processing steps
information
transfer to rejections
electroplating

Figure 5. Flow chart of the fully automated process chain

After polishing, the workpiece is transferred to the vision system to evaluate. The
workpiece will go to the next manufacturing step when no defects are found on the
surface.
A workpiece that has remediless defects will be rejected or discarded while the rest is
ground and polished according to the individual defect data. The retouched workpieces
are fed into the automatic vision system again until they are either accepted as qualified or
discarded as defectives. The compensation machining process for detected errors on the
surface is already realized in the system.

4. Simulation of Belt Grinding Processes

The operator can design the grinding and polishing path easily by using the software as
shown in Fig. 4. However, it is a highly experience-related job to decide the paths on the
workpiece surface. A skilled operator generates the paths and according programs that
result in a good surface quality and low rejection rate while an untrained operator reaches
a high rejection rate that damages the profitability of the robot system. The individual
operator produces inconsistent standard of design and causing a various surface quality of
the workpieces. Therefore, the belt grinding process and robot system should be studied in
detail with a process simulation being necessary at the same time. Both path planning and
parameter selection of the robot system will benefit from the results of the simulation
afterwards.

4.1 Introduction of Belt Grinding Processes

Belt grinding is a machining process with a geometric indeterminate cutting edge. The
grinding belt (cutting tool) consists of coated abrasives and is attached around at least two
rotating wheels.
The part to be ground is pressed onto one of these wheels, which is the contact wheel. Fig.
6. shows the belt grinding process. The material is cut off under non-permanent touch
between workpiece and abrasives.

760
 elastic grinding
wheel
grinding
belt workpiece

FA
VS

Figure 6. Belt grinding process with elastic wheel

The variant of the belt grinding with elastically deformable contact wheel is especially
suitable for the finishing of free-form surfaces because the elastic contact wheels allow a
flexible processing through their adaptation to the part surface. Compared to other
finishing processes, belt grinding with elastic contact wheel is characterized by a higher
removal rate with comparable surface quality. The most important advantages of this
process variant are:
- The elasticity of the contact wheel leads to a better form adaptation to the surface of
the workpiece compared to a rigid contact wheel and therefore produces a higher
quality of sculptured surfaces.
- The compensation of the infeed and orientation error of workpieces or tools is
possible to some degree.
- During a long phase the grinding belt wear factor is approximately constant. Thus
the belt wear does not affect the machining process in this phase.

4.2 Removals in Belt Grinding Processes

The most important point in simulation is to calculate the removals from the workpiece
surface at a discrete point of time. The simulation of the grinding process is more difficult
than those precise operating processes, e.g. milling, because the removal of the surface
cannot be obtained directly by calculating the swept volume through CAD data and tool
shape. The cutting tool in belt grinding is the grinding belt that is coated by hard
abrasives. In machining, the grinding wheel rotates and the belt rubs and strikes the
workpiece surface.
Because shape and distribution of the abrasives on the belt are non-uniform and rather
disordered, the belt grinding processes are considered as a cutting processes with an
indeterminate cutting edge. Another problem is the elasticity of the tool (grinding wheel)
that can cause a strong force variation between the grinding wheel and the workpiece.
Calculating removals is therefore not a purely geometric computation, but a quite
experience-based process. Besides the material of grinding belt, quite several other
parameters can simultaneously work on the final removal from the workpiece surface, for
example elasticity of grinding wheel, temperature and so on. The removal from the
workpiece surface is a function of a group of factors.
To model this, Hammann presented a linear experiential formula (Hammann, 1998;
Schueppstuhl, 2003):

761
 Vb
r = K A .kt . .FA (1)
Vw .lw
where γ is the removal; KA, a combination constant of some static parameters; kt, the
grinding belt wear factor; Vb, the grinding velocity; Vw, the workpiece moving velocity; lw,
the length of the grinding area; FA, the acting force. Using this model, one can do many
experiments making only one factor variable and all other factors unchanged at the same
time. Then, the experimental results can be combined to determine the coefficients in the
model. The parameters in the model are all one-valued which means all inputs of the
model are indicated by only one value, for example the acting force FA is the global force
between the workpiece and grinding wheel. The global grinding model is sufficient to the
operating of simple shapes. But it is obvious that they are incapable to free-form surface
grinding because not only the total removal but the local removals distribution (removals
at small sub-area) are neccessary to be known.
Removals distribution results from the contact force distribution in the grinding process.
The detailed local force distribution information (not only the global force FA) should be
obtained before the removals distributions are considered. Normally, the Finite Element
Method (FEM) is the traditional way to calculate the contact forces according to the initial
contact situation. The FEM deals with this contact problem as an ideal Signorini contact
problem (Krause, 2001) which solves an optimization problem by a contact energy
minimization principle. Considering the process locally, the surface of the treated
workpiece is split into a number of "finite elements" (Hammann, 1998; Schueppstuhl,
2003), assuming constant distribution of the contact pressure and cutting speed at each
element. Hence, the contact pressure must be calculated for each element. The overall
contact pressure distribution for one contact situation is then given by all local element
pressures. Local removals can be calculated through a multi-factor statistical analysis in
the second modeling step. Fig. 7. illustrates the framework of the process model for free-
form surface grinding. First of all, the geometry and elasticity information are constructed
as the initial contact situation which is the input of the FEM model. Then, FEM works out
the distribution of force between the workpiece and the grinding wheel. Finally, the
distribution of force together with other process parameters are given into the local
multifactor analysis model to calculate removals distribution.

contact situation
geometric
information Free-Formed Surface
Grinding Process Model
force distribution
by FEM

elasticity information experimental

plan removal
distribution
multivariate
analysis
grinding time, rotating velocity, belt material, workpiece material,
local relation
......

Figure 7. Process model for free-formed surface

762
4.3 Simulation of Robot Controlled Belt Grinding Processes
Fig. 8. is a flow chart of simulation processes (Schueppstuhl, 2003; Kreis& Kneupner,
2001a, 2001b). The planned path is known at first and the grinding process is divided into
some discrete time intervals. At any time point, the initial contact situation can be reached
by the current CAD model and the path. The next step is to calculate the force distribution
and then get the removals distribution by the process model. The current CAD model is
updated and moves to the next time interval until the path end. In this simulation
framework there are four important parts:
- Initial contact situation modeling
- Force distribution calculating
- Process model
- Workpiece model

Additionally, a direct control system should be implemented if an on-line control of the

robot is required (Kneupner, 2002; 2004).

local force
distribution
local contact
process
representation
parameters
path begin
local removals
simulation circle distribution
path end

updated CAD model

CAD model
next step

Figure 8. Simulation circulation

While simulating, the removal is calculated by the process model. Especially for surfaces
with a small radius, the change of the surface can be dramatic. Therefore it is not possible
to calculate a complex geometry using a so-called swept volume which interpolates the
removal between two different removal calculations. Instead, the shape of the workpiece
is changed directly after one removal calculation and the next calculation is based on the
new geometry. An important effect which should be taken into account in simulation
processes is the so-called “cut loose” effect (Kneupner, 2004). It means that the tool will be
not in contact with workpiece any more after a specific time of grinding if the workpiece is
not moved because all material in between will be worn out. While this happens, the
actual removal rate will slow down. It is clear that such a process can only be simulated if
the workpiece model is updated after a removal calculation. Thus, the time interval should
be small enough to neglect the decreasing removal rate because of the “cut loose” effect.
Moreover, a fast calculation is essential for practical reasons. The cycle time is driven by an
external cycle time. As it is possible to calculate the position of the tool in relation to the
workpiece with a high accuracy within the robot interpolation cycle time, we use a
multiple interpolation time cycle. This is an amount of time small enough for neglecting
the shape-change within a calculation. The Height Model is put forward in our project to
describe the initial contact situation specially considering the characteristics of belt
grinding process, assuming that workpieces for grinding are idealistically hard and stiff
without deformation and that the grinding wheel is made of soft material with a known

763
elasticity, which at the same time is also a prerequisite of the Signorini contact problem. So
when the grinding wheel contacts the workpiece, it deforms according to the geometry of
the workpiece surface and actual infeed. To describe such a situation, Height Model
divides the contact area into a mesh where then the initial contact situation of grinding
wheel can be encoded as a group of so-called Heights, which are intervals between the
base plane and deformed surface of the grinding wheel at all mesh points (see the right
part of Fig. 9.). The base plane is always vertical to the infeed direction and has invariable
distance to the wheel center. The normal vectors of all contact points on the deformed
surface are also recorded for later use.

Figure 9. Height Model

The Height Model is actually a discrete description of an initial contact situation in which
mesh size is a control factor for different precision requirement. In this way, each contact
situation can be described by a heights matrix H, which has the following form:

⎛
h
⎜ 11 h12 … h1n ⎞⎟
⎜ ⎟
h ⎜ h22 … h2 n ⎟⎟
⎜ 21
H= ⎜ ⎟
M⎜
⎜
M O M ⎟⎟
⎜ ⎟
h ⎜ m1
⎝
hm 2 … hmn ⎟⎠

where m, n is the mesh size in two directions.

The contact situation between workpiece and tool can change very fast during the
processing of sculptured surfaces, causing a strong removal variation. This fact asks for
quick calculation of force distribution in real time. The traditional way to calculate the
force distribution is resorting to the FEM that consider the contact problem as an ideal
Signorini problem. Blum and Suttmeier (Blum et al. 2000; 2003; Suttmeier, 2001) worked
out a FEM model. The FEM model has to achieve the optimization solution through
iterating steps each time when a new contact situation is presented. Thus, it is very time-
consuming. Although the model used an optimized mesh discretization, it still requires
about 15 minutes for calculating one contact situation. This is far away from the demand
for real time simulation. The force distribution calculation is the practical bottleneck of the
real-time simulation. To meet this, a new force distribution calculation model (Zhang et al.,
2005a; 2005b) is worked out to accelerate the calculation. The idea is that the neural
network or Support Vector Machine serves as an approximation of the FEM model to learn
the non-linear mapping from the initial contact situation to force distribution. The
expensive iteration process is shifted in the training process of the model, but not in run
time resulting in a dramatic reduction of the calculating time which makes the real time
simulation of the grinding processes possible.

764
After the local forces are known, it is time to decide the parameters in the process model.
The experiments and previous research (Koenig, 1996; Meyerhoff, 1998; Hammann, 1998;
Schueppstuhl, 2003) showed that an overall mathematical description of the belt grinding
process is not possible because a complete list of the influential factors cannot be
determined exactly.
In the process model, only eight parameters are selected out which are:
1. force floc (N/mm2)
2. rotating rate of grinding belt vb(m/s)
3. grinding time ts (s)
4. local radius of the workpiece rloc (mm)
5. the grain size of grinding belt kb
6. belt tension fb (N/mm2)
7. contact length lc (mm)
8. material of the workpiece mw

The influential factors of belt grinding can be determined only statistically due to the
geometrically indeterminate cutting edge. The parameter values must be in a reasonable
range in order to predict the removal precisely enough.
Statistical design of experiments is applied for the modeling of local relations. An essential
aspect of the statistical design of experiments is the fact that several factors are varied
simultaneously from a single experiment to the next one. In order to implement that, a
experimental plan is used by the statistical design of experiment. With the help of such an
experimental plan, more complex relations can be modeled. In our case a full factorial
experimental plan with 2 = 256 single experiments is used for all the 8 influential factors
8

on two factor levels.

The local removals as the response variables are measured after every individual
experiment using a sensor device. The experiment data are listed in the Table 1.
The influence of wear is minimized by the use of wear resistant grinding belts. A heat
influence can be minimized through a suitable selection of grinding times. The grinding
time for every individual experiment must not be too long so that the grinding belt is not
heated up. This experimental plan has been done only partially and is still in work.

floc vb ts rloc kb fb lc mw
2 2
(N/mm ) (m/s) (s) (mm) (N/mm ) (mm)
1 0.01 10 5 30 P60 0.1 5 1.0037
2 0.5 30 15 200 P120 0.4 50 brass
Table 1. The technical data for all parameters

4.4 Simulation Examples

Four parameters are used to simplify the definition of the workpiece surface, the infeed x,
the local curvature (the reciprocal of the radius) in y and z direction and the turning angle.
The simulation results can imply how this local parameters act on the local removals. Fig.
10(a). illustrates the simulation result with different infeeds. The length of the pillar on the
workpiece indicates the grinding infeed and colors on the workpiece surface denote the
local removal distribution. The infeed starts with zero at the beginning of the path and
reaches the maximum at the end of the path. The removal is growing when the infeed
increases. This is because a bigger infeed leads to a larger deformation of the elastic
grinding wheel and bigger acting forces. Fig. 10(b). illustrates the simulation result with

765
different local curvatures in y direction. The length of the pillar on the workpiece indicates
the local radius of the workpiece surface. The green pillar means the local surface is
concave and the red pillar means the local surface is convex. The infeeds are the same
along the grinding path. It shows that the convex part has a bigger removal than that of
the concave part which conforms to the practical experiments. Fig. 10(c). is the result with
different local curvatures in z direction. Fig. 10(d). is the simulation result of grinding
along a simple equal-infeed path. The workpiece has a small turning angle that changes
the local contact situation between the workpiece and grinding wheel, so the removals are
also not distributed uniformly.

(a) (b)

(c) (d)

Figure 10. Simulation results (a) with different infeeds, (b) with different longitude local radii, (c) with
different latitude local radius, (d) with turning angle (Schueppstuhl, 2003)

5. Defects Detection and Classification on Workpiece Surface

As mentioned in the section 3, another troublesome work is to find and identify the defects
on the surfaces that are caused in the previous casting phase. In the past, this job was done
by well-trained workers. They checked the surfaces for potential defects and classified
them into pre-defined categories. In order to reduce the wage cost, this process is also
attempted to be replaced by an automatic defect detection and classification with a vision
system.

defects feature classifier

defects location segmentation extraction

Figure 11. The framework of the automatic identification and classification system

766
The planned classification system mainly consists of four parts: defects location, defects
segmentation, feature extraction and classification. The four parts in the system act
sequentially. Fig. 11. shows the framework of the classification system.
At the beginning, the system detects and locates the defect in the grayscale bitmap
obtained from the vision system. Before encoding the bitmap into some meaningful
feature values, the area of the defect in the image should be pre-processed and segmented.
In this phase the separate defect image is obtained and is ready for feature extraction. The
feature extraction is the most important part in the system. It defines the rules of
describing and expressing the defects inside an image in a form that a classifier can
understand. At first, those features should be found which are useful to differentiate one
defect from another,.
The features are not limited to the shape feature, but can be the texture features and some
statistical features of the segmented image. Then, an effective way should be found to
encode such selected features, specially in a mathematical way. After the data pre-
processing, the feature data are applied as the training data to the classifier. Many artificial
intelligent techniques have the capability of multi-class classification, e.g. k-NN, MLP
network, RBF network, SVM, etc.
The vision system adopt a specific illumination method to get a light color surface lying on
the dark background. A small window is then applied on the grayscale picture of the
surface. By analyzing the grayscale distribution in each small window, the defects can be
located precisely because the defects are of deep color and surrounded by the light color
surface. Fig. 12. shows 5 different defects selected from the total 19 classes that have been
found on the workpiece surface.

Figure 12. Example of 5 classes of defects. (1)pore, (2)subtle line, (3)fluff, (4)polishing shade (5) casting peel

767
Technicians differentiate the defects by experience. Actually such experience involves
some rules. For example, a small black solid circle in the white background may be the
“pore” defect; a long stroke on the surface may be a “subtle line” defect. In other words, a
technician draws a conclusion by combining some attributes of the defect in the
segmented image, e.g. size, grey scale distribution and so on. All these attributes used by
technicians (no matter intentionally or subconsciously) to discriminate defects are called
valid features of the defect.
The length and width of the defects are the simplest valid features. It is very easy to see
that such size information are the reason for a technician to differentiate the defects.
However, not all features used by human beings are as straightforward as the shape
features such as length and width. The fact is that you use some features to classify but
you do not know you use them as a base. From the result of some other similar projects
(Lampinen&Smolander 1994; Iivarinen&Visa 1998), three kinds of features are normally
extracted to identify and classify the object, shape, texture and some statistical features of
the image, e.g. grayscale histogram. It is not obligatory to include all those features as the
input to the classification process. It depends on the project itself to decide which features
to include.
Two basic principles should be hold in mind:
1. The features extracted are sufficient to separate different defects.
2. The features that is not helpful to discriminate the defects are not extracted to the
classifier.
The first principle is a precondition of the system to function. Different defects are
represented by features of different values. Think of the classifying of an apple and a pear,
for example. Only by the shape feature it is not possible for people to tell whether it is an
apple or a pear. It does not comply with the first principle because the apple may have the
same and resemble the outer shape of the pear. The second principle is to ensure the
generalization of the classifier.
Those features that are not helpful to distinguish the objects are not selected best. For
example, the weight is not a feature that should be used because an apple is possibly
heavier or lighter than a pear. Unnecessary features act like noise in the system. They
distract the classifier from the right way and influence the performance of the classifier.
The quantitative measurements of features are called descriptors. The way to measure
shape are called shape descriptors while texture features are expressed by texture
descriptors. There are many different descriptors available. Descriptors design and
selection is an important research topic in the classification system. The current system
uses only some simple shape descriptors, e.g. length, breadth, compactness and roundness
of the defect. The right classification ratio is only about 30% now. There are three reasons
for the low classification ratio.
The first one is the lack of the effective descriptors to account for the defects. The second
reason is that the system did not use the up-to-date intelligent classifier and the last one is
that there are two many pre-defined classes of defects (19 classes). Some kinds of defects
are so similar that even skilled people can not differentiate them from each other. For
example, “polishing shade” is similar to the “subtle line”. When only analyzing the
grayscale pictures, the “pore” defect is difficult to separate from the “fluff” defect by the
operator in some cases, even though they are different in reality, see Fig. 12. The
classification of defects is still under development. It is supposed to be able to reach an
80% right classification ratio by well-presented feature extractors and tailored classifier in
the future.

768
6. Summary
Due to the tendency of the globalization of markets and competition from low wage
countries, it is profitable for the industry to further automate the robot-controlled grinding
and polishing system, especially in the sanitary fitting industry. The high automation level
of the robot system not only relieves the human being from laborious tasks, but also
elevates the efficiency of the whole manufacturing process by producing a steady product
surface quality. A framework of such a robot system is put forward in this paper. The
automation of the robot system focuses on two parts. One is to facilitate the grinding and
polishing paths planning processes. Software is developed to help the operator to generate
the grinding and polishing paths quickly and easily. It offsets the incapability of universal
off-line programming systems for grinding and polishing. With this new software the
tasks that had to be done by high-specialized staff before can now be taken over by normal
operators. To help the operator to design the paths, the belt grinding process and its
simulation are carefully studied. Another key point of the automation is to automatically
find and classify the defects on the workpiece surface into some pre-defined classes in
order to apply an according compensating machining process. The detection of the defects
is already realized while the classification should be improved in the future work.

This research and development project was funded by the German Federal Ministry for Education
and Research (BMBF) within the national "Research In Production For Tomorrow"
Programme and managed by the Project Agency for Production and Manufacturing Technologies
(PFT), Forschungszentrum Karlsruhe

7. References
Blum, H; Suttmeier, F.T. (2000) An adaptive finite element discretisation for a simplified
signorini problem. Calcolo, 37(2):65-77.
Blum, H; Schroeder, A; Suttmeier, F.T(2003). A Posterori Error Bounds for Finite Element
Schemes for a Model Friction Problem. Conference Proceedings of Simuation Aided
Offline Process Design and Optimization in Manufacturing Sculptured Surface, pp
39-47, Witten-Bommerholz, Germany.
Cabaravdic, M; Kneupner, K; Kuhlenkoetter, B. (2003) Methods for efficient optimisation
of robot supported grinding and polishing processes. International Conference on
„Trends in the Development of Machinery and Associated Technology“, Barcelona,
Spain.
Hammann, G.(1998) Modellierung des Abtragsverhaltens Elastischer Roboter-gefuehrter
Schleifwerkzeuge. PhD thesis, University Stuttgart.
Iivarinen J; Visa, A. (1998) An Adaptive Texture and Shape Based Defect Classification. In
the Proceedings of the 14th International Conference on Pattern Recognition, vol. 1,
pp. 117-122, Brisbane, Australia
Lampinen, J; Smolander, S. (1994) Wood defect recognition with self-organizing feature
selection. Intelligent Robots and Computer Vision XIII: Algorithms and Computer
Vision, Proc. SPIE 2353, pp. 385-395.
Kneupner, K. (2002) Directcontrol. Ein Programmierkonzept fuer Roboterzellen. In
Conference proceedings of Robotik, Ludwigs-burg, Germany.
Kneupner, K. (2004). Entwicklung eines Programmier und Steuerungskonzepts fuer
Robotersysteme auf der Basis eines Umwel-tmodells. PhD thesis, Dortmund
University.

769
 Koenig, W. (1996) Fertigungsverfahren Band 2 --- Schleifen, Honen, Laeppen. VDI Verlag.
Krause, R. H. (2001) Monotone Multigrid Methods for Signorini's Problem with Friction.
PhD thesis, Freie University Berlin.
Kreis, W; Kneupner, K. (2001a). Ein Simulationsgerechtes Prozessmodell fuer das Freiform
Bandschleifen. Dortmund Germany.
Kreis, W; Kneupner, K. (2001b) Simulation des Bandschleifprozesses. Frontiers in
Simulation, Simulationstechnik 15th. Symposium, pp 517-522,. Paderborn Germany.
Meyerhoff, M. (1998) NC-Programmierung fuer das kraftgesteuerte Bandschleifen von
Freiform-flachen. PhD thesis, University Hannover.
Schueppstuhl, T. (2003) Beitrag zum Bandschleifen komplexer Freiformgeometrien mit
dem Industrieroboter. PhD thesis, University Dortmund.
Suttmeier, F:T. (2001) Error Analysis for Finite Element Solutions of Variational
Inequalities. PhD thesis, Dortmund University.

770

View publication stats