Information-theoretic and statistical methods of failure log selection for improved diagnosis

Diagnosis of each failed part requires the failed data captured on the test equipment. However, due to memory limitations on the tester, one often cannot store all the failed data for every chip tested. Consequently, truncated failure logs are used instead of complete logs for each part. Such truncation of the failure logs can result in very long turn-around times for diagnosis because important failure points may be removed from the log. Subsequently, the accuracy and resolution of final diagnosis may suffer even after multiple iterations of diagnosis. In addition, the existing test response compaction techniques though good for testing, either adversely affect diagnosis or are highly sensitive to deviation from the chosen fault model. In this context, the industry needs dynamic selection of better failure logs that enhances diagnosis. In this paper, we propose a number of metrics based on information theory that may help in selecting failure logs dynamically for improving the accuracy and resolution of final diagnosis. We also report on the efficacy of these metrics through the results of our experiments.