Abstract
A goal of quality assurance is to exhibit superior defect detection performance while minimizing the costs of inspection. In wire bonding, the acceptability of lots is decided based upon destructive test results of one or more samples drawn from the lot. To increase reliability of the inspection, it is necessary to conduct many destructive tests. Performance and costs are in a trade-off relationship, and thus it is desired to develop a method capable of improving the determining accuracy while preventing an increase in costs. For this reason, we focus on effective use of information obtained from "sample" as well as from "uninspected items". The former is accurate, but provides a relatively small dataset. Conversely, the latter is less accurate, but has an abundant dataset. The research described in this paper is an effort to improve classification performance of sampling inspection using the two concepts in a complementary way. In order to achieve highly accurate discrimination, we apply "Taming" - a machine learning problem - to the sampling inspection problem. Furthermore, sensing technology using a thin film Acoustic Emission (AE) sensor is implemented to evaluate the bonding process. The determining accuracy of the proposed method is compared with basic sampling inspection through actual bonding experiments.
