Predicting system-level test and in-field customer failures using data mining

Feature engineering with canonical analysis for effective statistical tests screening test escapes

“…It has been shown that some test escapes differ from normal chips in their test measurement values relative to the measurement means [2]. Fig.…”

“…Pixels that are spatially close but have significant difference in intensity will have smaller weights, while pixels that are a bit farther apart but very similar in Paper 24. 2 INTERNATIONAL TEST CONFERENCE 2 intensity will have larger weights. Therefore, the sharp edges of an image can be preserved and the noises are more likely to be filtered.…”

“…With the application of canonical transform, the useful information in separating the test escapes and the normal population in all generated features Paper 24. 2 INTERNATIONAL TEST CONFERENCE 5 is incorporated into the comprehensive test. An experimental comparison on the two exemplar flows will be made in Section V. Given any training set, it is possible that the manufacturing process drifts over time such that the training set is no longer sufficiently representative for the later data.…”

“…Each field return found at the customer side incurs significant cost and requires thorough analysis of the cause. It has been shown that a good fraction of field returns are test escapes that pass the complete test program, but fail at system level due to their intrinsic defects [1], [2]. Applying system level tests to each chip prior to shipment, however, is undesired because it often results in high test time and cost.…”

“…Butler et al [6] successfully demonstrated burnin minimization by a collection of multivariate analysis. Chen et al [2] showed various data mining techniques on final test data to predict system level test (SLT) failures.…”