微電子製程尺寸逐年微縮，致使產品產生多元的失效機制，同時失效機制亦是影響其壽命的重要因素。微電子資料常具有早夭、多峰、厚尾等特徵，且因試驗時間的限制下，亦可能產生設限資料。因此分析微電子資料頗具挑戰性，特別是當試驗樣本少時更為不易。實務上，混合模型與串聯系統模型常用以配適多重失效模式的資料，但兩模型間的本質差異卻不易區別。另一方面，工程師和試驗操作人員的專業背景與經驗，卻可以對模型提供重要的資訊。本文以獨立之對數常態分布配適各失效模式之資料，分別以混合模型和串聯系統模型剖析具多重設限之產品失效機制與可靠度分析。藉由隱藏變數與設限樣本之補值，於貝氏架構中以共軛先驗分布建立共軛結構，進而簡化計算的複雜度。將兩模型分別應用於閘極氧化層資料與電遷移資料中，顯示融合先驗資訊之貝氏方法可提供實用的可靠度分析，並可驗證專家建議之模型的適切性。;The year-by-year downscaling in microelectronics results in multiple failure modes which makes an essential impact on the products’ lifetime. Features such as infant mortality, multiple failure modes, and heavy tails are common in microelectronics data. Under the limitation of experimental time, it also yields censored data. Therefore, it is a challenging task for analyzing the microelectronics data, especially with a limited amount of test units. The mixture model and the series system model are often used practically. But the essential difference between these two models is hardly distinguished. On the other hand, experts with domain knowledge and experience can provide important information for the model. In this thesis, independent log-normal distributions are considered for the failure times of different failure modes in the mixture model and the series system model which are used to analyze the failure mechanisms and reliability of multiply censored products. By imputing the latent variables and the censored observations, a Bayesian modeling with conjugate priors is constructed to simplify the computation. The proposed method is applied to gate oxide data and elctromigration data. It turns out that the Bayesian approach incorporated with the prior information provides useful reliability analysis, especially in confirming the validity of the models recommended by the experts.
