Advanced Methods for the Analysis of Semiconductor Manufacturing Process Data - Advanced Techniques in Knowledge Discovery and Data Mining - page 61

Error due
to loose
label
Fig. 2.19. Feature space for vision system of medical laboratory robot.
selected object for each class from the underlying database. Thus, occurring
problems, e.g., misclassifications, and underlying causes, can be easily made
overt. This alleviates troubleshooting in system design and increases design
speed, reliability, and overall productivity. The work is extended to micro-
electronic manufacturing process data analysis and the features elaborated
in prior research and application projects are adapted to this domain. For
instance, data entries can be tracked back from the projection in the process
database as illustrated in Fig. 2.19 for image data. Thus, the database can
be browsed and analyzed according to the inherent clustering and structure
in the data. The extension of the existing approach to semiconductor manu-
facturing will be presented in the following section and in Section 2.5, giving
an outline of the envisioned domain-specific system.
2.4 Experiments and Results
The first step of the work in this feasibility study targets the validation
and demonstration of the actual practical assistance of the dimensionality
reduction and visualization approach to discover structure in and extract
knowledge from the industrial high-dimensional database. Thus, it is expected
from the visualization that the known split information can be effortlessly
retrieved from the map. In this case, unknown clustering in the data, due to
detrimental and unintended effects, could also be made overt to the process
analyst at a glance.
The most simple and fast Visor projection method was applied to the
data first [2.24]. Figure 2.20 shows that distinct yet overlapping clusters can
be identified in the data. It is well known from physical and technological
background knowledge, that the generated split affects only a fraction of