|
Redundant Test Identification
Hongfei Wang |
Shawn Blanton |
Reducing the number of specification-based tests for integrated, heterogeneous systems (e.g., phase-lock loops (PLLs), high-speed serializers/deserializers (HSSs), micro-electromechanical systems (MEMS), etc.) while maintaining product quality with low defect escape and yield loss is an important issue in test economics. This is especially true when some tests require complex setup procedures and expensive test execution. In previous research here at Carnegie Mellon, a new method that employs Boolean minimization has been developed to identify redundant tests, and has been successfully applied to fabricated HSS and PLL circuits from IBM. Boolean minimization is used to identify the relation between necessary tests and those found to be redundant, using sum-of-products representations. However, the first step in this method is to select a set of potentially-redundant test candidates, which is now carried out by brute force, i.e., checking the potential redundancy of each individual test one by one using all the other tests. This drawback leads to significant computation when the total number of tests is large.
In this work, a statistical learning technique, namely decision trees, is employed to provide an efficient and systematic approach for redundant-test candidate identification. Using the test data collected from HSS failing circuits, the resultant decision tree shown in Fig. 1 identifies 29 tests as potentially redundant. Further, in-depth analyses revealed that 14 of these 29 candidates did proved to be redundant. Importantly, the identification of the redundant test candidates reduced the computation needed by the follow-on, computationally intensive analyses by more than 27%. Since only binary test data is used in this work, this statistical learning technique can be extended to similar test-compaction scenarios, including the functional test of digital circuits.
|
