Advanced Methods for the Analysis of Semiconductor Manufacturing Process Data - Advanced Techniques in Knowledge Discovery and Data Mining - page 68

Fig. 2.25. Visualization of selected SPLIT6 novelty classification.
filtering within an information-processing hierarchy for process analysis, con-
trol, and optimization. The achieved results are illustrated in Fig. 2.25. The
complete training set itself was correctly classified with regard to the bifur-
cation normal (class 1) or novel (class 2). For the test set, the vectors of
classes 2 and 3 were also correctly identified as novel. However, numerous
vectors of the normal test data were also classified as novel, as they occur a
significant distance from the normal training data. Thus, a recognition rate of
only 87.2% was achieved for the test set. It must be minded that the superior
result of the RNN classifier required training by 95 vectors. The majority
of these samples were counterexamples from the abnormal or novel range.
In contrast, OCC was trained with only 30 vectors. The presented training
data are rather sparse, so improvements of the OCC performance can be ex-
pected by providing larger data sets of normal process data as well as by a
more sophisticated R max computation and resulting normal range coverage
in parameter space. However, though numerous practical improvements are
possible, the feasibility of the described method to filter out significant novel
data and perform as a data-reduction module also has been demonstrated.
The objectives of this feasibility study for the chosen problem and data
have all been achieved. The feasibility of the selected soft-computing methods
could be confirmed and relevant approaches for method improvement could
be identified.
2.5 Proposed System Architecture
In the presented feasibility study, several selected methods were investigated
with actual problem data with regard to their applicability for semiconduc-
tor manufacturing. As encouraging results have been obtained, a more so-