對於高可靠度 (highly reliable) 產品而言~，由於傳統的加速壽命試驗
(accelerated life test~，簡稱 ALT)~，無法在合理時間內~，收集足夠
的壽命資料~，因此~，衰變試驗 (degradation test)已成為廣泛地用來
評估高可靠度產品的可行方法~。但對於非常高可靠度 (very-highly
reliable) 的產品而言，由於其衰變速度極為緩慢，以致於無法於一合理
的時間內，進行有效的壽命推估，在此情形下，可採用較高的應力來加速
其衰變速度，再利用合適的壽命-應力物理模式(life-stress relation)
，外插以推估出在正常使用狀態下之可靠度資訊，此即所謂加速衰變試驗
(accelerated degradation test~，簡稱 ADT)。在設計與執行一衰變試
驗時~，常面臨「試驗進行中」和「試驗進行前」~，兩類之設計問題~，
其中要考慮的重要決策變數包括：在各試驗應力水準下，樣本之個數~、
試驗時間之長短以及觀測值之量測頻率等等~。這些變數對試驗費用與資
料分析之精確度影響甚鉅~，因此~，本論文主要即探討衰變試驗之設計~
，以提昇資料分析之精確度。其中~，我們提出一直觀且即時 (on-line)
之決策準則~，以協助實驗人員於實驗進行中~，即能決定一適當之試驗終
止時間~。另外，針對實驗進行前之準備~，我們建構一非線性之整數規劃
模型~，在試驗費用不超過預算 (budget) 的前提下~，最小化產品在正常
使用條件 (normal use condition) 下~，產品壽命分配之 $(100p)th$
百分位數和平均值 (mean-time-to-failure~，簡稱 MTTF)估計值的變異
數~，以求出非加速衰變試驗 (non-ADT) 與 ADT 之各試驗應力水準下~，
最適之樣本數~、終止時間及量測頻率~。

For highly reliable products, traditional accelerated life-test
(ALT)methods may not be practical to estimate the lifetimessince
the products are not likely to fail in a reasonable amount of
time.Hence, degradation tests are widely used to assess
thereliability of highly reliable products.However, for very-
highly reliable products, the degradation rate is too slowto
make useful inferences within a reasonable amount of time. In
this case,higher stresses are usually used to accelerate the
degradation rate anda suitable life-stress model is adopted to
extrapolate and estimate theproduct''s reliability at a design
stress. This method is called theaccelerated degradation test
(ADT).In designing an ADT, there are many important decision
variablessuch as the sample size, the length of a life-
testingand the inspection frequency at each stress level. These
variablesare influential tothe cost and the precision of data
analysis.Therefore, this thesis focuses on designing an ADT to
enhance the precisionof data analysis. An intuitive on-line
procedure isproposed to determine an appropriate termination
time for ADTs.Furthermore, by using the criterion of minimizing
the variance of estimatingthe $(100p)th$ percentile of the
product''s lifetime distributionand the product''s mean-time-to-
failure (MTTF),a nonlinear integer programming is established to
determine the optimal(a)the sample size to test, (b) the length
of a life-testing, and (c)the inspection frequency at each
stress level.