Industrial processes often have to separate
good parts from bad parts on the basis of qualitative criteria. One
has to select a criterion and then choose a threshold such that all the
parts above the threshold will be classified as "bad", and all the parts
below the threshold will be classified as "good". However, such
inspection tasks -- flaw detection, for example -- are difficult to
automate because:
- A good criterion
may need to be a combination of several measurements. - It can be
difficult to quantify inspection tasks that require some level of
"judgment". - Even with a good criterion one still must find the
optimal value for the threshold.
The ACG Linear Classifier emulates some
components of human judgment by combining criteria. The classifier
computes an optimal linear combination of measurements so as to separate
clearly the "good" and "bad" parts. To use the classifier the user sets up
inspections to measure various part parameters (e.g. gray level, size,
etc..) and then trains a representative sample of parts, telling the
system whether each sample is good or bad. The classifier will then
automatically determine the best combination of inspections, and suggest a
threshold for pass/fail. Advanced features allow the user to view the
training data on a 3-D plot to check the separation of
data. |