Bibliographic Metadata

Title
Application of bayesian networks and stochastic models to predict smart power switch lifetime / Katrin Plankensteiner
Additional Titles
Application of Bayesian networks and stochastic models to predict smart power switch lifetime
AuthorPlankensteiner, Kathrin
CensorPilz, Jürgen ; Wagner, Helga
Published2015
DescriptionXX, 156 S. : graph. Darst.
Institutional NoteKlagenfurt, Alpen-Adria-Univ., Diss., 2015
Annotation
Abweichender Titel laut Übersetzung der Verfasserin/des Verfassers
Zsfassung in dt. Sprache
LanguageEnglish
Bibl. ReferenceOeBB
Document typeDissertation (PhD)
Keywords (DE)Zuverlässigkeit / Leistungshalbleiter / statistische Lebensdauermodellierung / Bayes / Vorhersage
Keywords (EN)Reliability / Smart power semiconductors / Statistical lifetime modeling / Bayes / Prediction
URNurn:nbn:at:at-ubk:1-26351 Persistent Identifier (URN)
Restriction-Information
 The work is publicly available
Files
Application of bayesian networks and stochastic models to predict smart power switch lifetime [11.27 mb]
Links
Reference
Classification
Abstract (English)

Testing the reliability of Smart Power semiconductor devices is highly time and cost consuming. Nevertheless, it is substantial, since in automotive applications semiconductor devices are used for passenger safety, e.g. for airbags. To save test resources, commonly accelerated stress tests in combination with statistical models are applied to achieve reliable predictions for the device lifetime. For this purpose, the development of a valid lifetime model is the aim of this thesis. For the analysis of lifetime data, two regression models are presented: a linear regression method via a Bayesian Network model and a generalized regression method using a Gaussian Process prior. A main challenge is the highly complex data following a mixture distribution with two components representing two different physical failure mechanisms. Moreover, data shows censoring. In the first part of the thesis, the Bayesian application of the Bayesian Network model is analyzed. Since the amount of data is assumed to be too small for an efficient parameter estimation of a full model, different submodels based on a preselection with physical knowledge are investigated and compared. In the second part of the thesis, the Bayesian application of a standard Gaussian Process regression model is evaluated. For the covariate selection, the squared exponential covariance function, which utilizes an automatic relevance determination for each covariate in the model, is applied. The selection of an appropriate covariance function is based on physical knowledge. Using the most significant covariates for the mean lifetime, the model is extended with a model for the standard deviation. For both regression methods, the Bayesian Information Criterion and the Bayesian Factor are used for model comparison. To evaluate the prediction quality, cross-validation, posterior predictive distributions, and Sum of Squared Errors of Prediction are applied. The results show that the obtained lifetime model based on the whole data set leads to accurate predictions, whereas a lifetime model based on a data subsample does not lead to reliable forecasts. Finally, ideas for stress profile modeling using damage path models are presented. For this purpose, single damage paths are accumulated to obtain the damage path of cumulated stress conditions. Empirical and statistical results are discussed and it is shown that Miner's rule for linear damage accumulation can be applied with a safety factor.

Stats
The PDF-Document has been downloaded 25 times.