Understanding customer returns from a test perspective

Tikkanen

2012 IEEE International Test Conference

et al. 2012

Self Cite

This work studies the potential of capturing customer returns with models constructed based on multivariate analysis of parametric wafer sort test measurements. In such an analysis, subsets of tests are selected to build models for making pass/fail decisions. Two approaches are considered. A preemptive approach selects correlated tests to construct multivariate test models to screen out outliers. This approach does not rely on known customer returns. In contrast, a reactive approach selects tests relevant to a given customer return and builds an outlier model specific to the return. This model is applied to capture future parts similar to the return. The study is based on test data collected over roughly 16 months of production for a high-quality SoC sold to the automotive market. The data consists of 62 customer returns belonging to 52 lots. The study shows that each approach can capture returns not captured by the other. With both approaches, the study shows that multivariate test analysis can have a significant impact on reducing customer return rates especially during the later period of the production.

Section: Prior Related Workmentioning

confidence: 99%

Screening customer returns with multivariate test analysis

Tikkanen

2012 IEEE International Test Conference

et al. 2012

Self Cite

“…In another example, the authors in [11] analyzed parametric wafer sort data from a high quality SoC and showed the potential for building models from the test data which were capable of predicting devices likely-to-fail at final package testing. Similarly, multivariate test analysis was used in [10,12] to predict parts that would fail in the field, i.e. customer returns.…”

Section: Related Workmentioning

confidence: 99%

“…Test selection algorithms were applied in previous work to identify the tests which are most important in describing the failing signature for customer returns [10,11,12]. In this work, we apply the same SVM test ranking algorithm, which applies the C-Support Vector Classification (C-SVC) algorithm [1] to determine the importance of each test.…”

Section: Test Selection For Multivariate Test Analysismentioning

confidence: 99%

An experiment of burn-in time reduction based on parametric test analysis

2012 IEEE International Test Conference

Abadir

2012

Self Cite

Burn-in is a common test approach to screen out unreliable parts. The cost of burn-in can be significant due to long burn-in periods and expensive equipment. This work studies the potential of using parametric test data to reduce the time of burn-in. The experiment focuses on developing parametric test models based on test data collected after 10 hours of burn-in to predict parts likely-to-fail after 24 and 48 hours of burn-in. Our study shows that 24-hour and 48-hour burn-in failures behave abnormally in multivariate parametric test spaces after 10 hours of burn-in. Hence, it is possible to develop multivariate test models to identify these likely-to-fail parts early in a burn-in cycle. This study is carried out on 8 lots of test data from a burn-in experiment based on a 3-axis accelerometer design. The study shows that after 10 hours of burn-in, it is possible to identify a large portion of all parts that do not require longer burn-in time, potentially providing significant cost saving.

“…For predicting customer returns, the works in [7], [8] apply test selection when learning from an extremely imbalanced dataset, where the imbalance was on the order of 10k vs. 1, i.e. 10k passing parts versus 1 failing part.…”

Section: Related Workmentioning

confidence: 99%

Forward prediction based on wafer sort data — A case study

Drmanac

2011 IEEE International Test Conference

et al. 2011

Self Cite

This paper studies the potential of using wafer probe tests to predict the outcome of future tests. The study is carried out using test data based on an SoC design for the automotive market. Given a set of known failing parts, there are two possible approaches to learn. First a single binary classification model can be learned to model all failing parts. We show that this approach can be effective if the failing parts are compatible in learning. Second, an individual outlier model can be learned for each failing part. We show that this approach is suitable for learning failing parts such as customer returns, where each may have a unique failing behavior. We also show that with Principal Component Analysis (PCA), a learning model can be visualized in two or three dimensional PC space, which facilitates an engineer to manually select or adjust the model.