The cost of poor quality in the steel industry is continuously
increasing. To address one aspect of this problem in the manufacture
of flat products, an Italian research centre has devised a prototype
system for the real-time inspection of stainless steel strip. Poor
quality in this industry can lead to costly in-plant repairing, product
downgrading and the risk of claims, particularly in light of the present
laws concerning quality assurance, which comes back to the supplier.
A digital image of the strip obtained on the production line is analysed
for defects which are then classified into at least six types by one
of three image recognition procedures. Applications are foreseen in
a wide range of processes involving the manufacture of continuous
material of uniform surface appearance, such as in the sheet metal,
papermaking and plastics industries.
Many manufacturing processes produce continuous
strips of material - steel and paper making are examples - in which
the product would be spoiled, and even rendered worthless, by surface
blemishes occurring during manufacture. Centro Sviluppo Materiali
(CSM), a materials research centre based in Rome, has long been
investigating methods of automating the surface inspection of steel
strip so that faults can be identified swiftly and efficiently.
In an earlier steel research project, CSM collaborated with steel
maker AST to devise an automated method of examining stainless steel
strip emerging from the annealing and pickling line at the Terni
steelworks in Umbria. That pilot project, completed in February
1994, was so successful that CSM was awarded further funding to
develop a prototype suitable for commercial application.
The prototype can inspect and analyse a complete steel strip as
fast as it can be manufactured. It consists of two main parts, the
Remote Acquisition Module and the Surface Inspection Module.
The Remote Acquisition Module is mounted close to the production
line. It contains a special camera which scans a line of 2,048 points
across the width of the strip producing an image in 256 shades of
grey. As the strip passes the camera, successive lines are added
to build up a complete image. The resolution of the image depends
on the size and speed of the steel strip, but is typically half
a millimetre to one millimetre.
If required, a second camera may be used to view the strip from
a different angle, as this usually makes it easier to detect the
complete range of defects.
The image is corrected for non-uniform illumination within the module
and then the digitised data is sent by a fibre-optic link to the
second major component, the Surface Inspection Module. Because the
amount of image processing required here is huge, much of the computation
in the module is carried out on hard-wired boards rather than by
a program held in memory. This is the only way the surface inspection
can keep pace with the speed of strip production.
As the manufactured surface is designed to be of uniform appearance,
defective areas generally show up where points in the image ('pixels')
differ in brightness from their surroundings. The first stage of
processing is a map of the strip showing individual defective pixels.
Next, neighbouring defective pixels are grouped into objects whose
'features' which are defined by 14 parameters including their geometric
properties (dimensions, area and perimeter) and their optical characteristics
(such as the maximum and minimum brightness). As many as 1,024 defects
can be handled at once. The parameters of each feature are then
passed to a classification unit which uses three different methods
to decide what kind of defect is present. For steel strip, for example,
the defects could be scratches, dents, bumps, stains, manufacturing
flaws and so on. Defects arising from impurities are particularly
important to recognise because most stainless steel products are
sold mainly by appearance, rather than their mechanical properties.
(For example, cutlery, pots and trays.) In addition, impurities
can break the metallic structure and create a weak point mechanically,
or they could affect the rust-free feature of stainless steel.
The first classification method, widely used in image-recognition
applications, simply compares the set of parameters with a standard
set for each type of defect to determine which class they fall into.
It is not very efficient in distinguishing defects where the permitted
ranges of parameters may overlap with each other.
Another method, one of several possible statistical approaches,
compares each defect with a sample of defects from a database and
calculates the probability of a match to each one. The class with
the highest probability is chosen. This method achieves 70-75% correct
A third method uses a neural network, a simple computer which can
be trained to recognise different kinds of defect by being shown
many examples. Although its success rate is only slightly better
than the statistical method, it works much faster.
All three methods are available to the operator, who can view the
results displayed on a standard PC and monitor. The final result
is a map of the whole strip, showing the location of each defect
marked by a symbol according to its type.
The system has been tested on a large database of known defects,
and is now 75-80% successful in distinguishing six classes of defect
on stainless steel strip. CSM is looking to improve this figure
by 'context analysis' which examines the relationship between each
defect and neighbouring defects. The classification efficiency would
need to reach 90% or higher to make it commercially attractive.
Still a role for humans
The original intention of a completely automated inspection system,
however, has been modified in the light of experience. It will not
be practicable in the near future to do away with visual inspection,
because it is not yet possible to completely replace human skill
and experience in recognising the many forms of defects in the steel.
How useful would such a system be in practice? While defects on
steel strip cannot normally be remedied, the inspection system does
provide rapid warning of problems in the manufacturing process which
can be dealt with before too much steel is wasted. It would be particularly
valuable at early stages of manufacture, where faults can be identified
before the material is passed to subsequent stages. In addition,
automatic inspection of the finished strip could be combined, for
example, with selective cutting, so that only the best sections
of strip are used and defective portions rejected.
CSM envisages applications for the inspection system in any industrial
process in which sheet material of uniform appearance is produced
in continuous strips, such as papermaking, plastics and metals other
than steel. Of course, the nature of the defects will be different
from process to process, but the principle remains the same.
CSM is already negotiating with a potential partner to commercialise
the system. There are several on-line inspection systems on the
market, but none is efficient at classifying defects correctly.
The classification algorithm is the most innovative component of
the CSM system and could even find market potential in its own right.