高可靠度產品之衰變試驗分析

Analysis of Degradation Tests for Highly Reliable Products

彭健育(Chien-Yu Peng)

指導教授 ： 曾勝滄 ； 鄭少為

國立清華大學/理學院/統計學研究所/博士(2008年)

https://doi.org/10.6843/NTHU.2008.00206

摘要

衰變模型目前已廣泛地被用來評估高可靠度產品的壽命資訊。當產品存在一與壽命具高度相關之品質特徵值(Quality Characteristics, QC) 時, 則可透過此品質特徵值的衰變量來估計產品壽命分配, 因此, 評估產品壽命的準確度與產品衰變路徑的建模工作有著極密切的關連性。本文將針對衰變試驗所面臨的建模及相關資料分析進行研究。
(i) 文中首先提出一般化線性衰變模型, 其優點是可以同時將測試產品之間的差異性、衰變量與時間的關聯性及量測誤差等因素納入考慮。在此衰變模型下, 可推導出產品壽命之分配, 並進一步探討當衰變模型(隨機效應模型及Wiener 過程模型) 之間發生誤判時, 對產品平均壽命(mean-time-to-failure, MTTF) 之影響。一般而言, 在大樣本的假設下, 衰變模型之間的誤判對MTTF 之影響並不嚴重, 然而, 在小樣本或測試時間較短時, 其誤判的影響程度將不容忽略。
(ii) 當試驗總成本受限時, 本文採用平均壽命估計量之變異數極小化為準則, 提出一簡便演算法來決定執行前述衰變試驗所需之最適樣本數、測試頻率以及量測次數, 進而可精確地推估產品可靠度資訊。最後, 藉由敏感度(sensitivity) 分析,可探討衰變模型中之參數及試驗成本發生變動時, 對最佳實驗配置之影響。
(iii) 從實證資料可發現, 衰變模型中單位時間的衰變率(mean degradation rate)大多為正值且其分配並非左右對稱, 故採用偏斜常態(skew-normal) 分配將更適合用來描述產品的衰變路徑。在隨機效應為偏斜常態分配之假設下, 本文利用EM (Expectation-Maximization) 演算法來估計模型中之參數, 並推導出產品
之壽命分配。此外, 文中亦探討模型選擇(model selection) 問題, 由laser 的實證資料, 可發現此衰變模型較傳統常態假設下的隨機效應模型更具穩健性(robustness)。
(iv) 在新產品研發階段, 欲進行可靠度壽命測試時, 通常僅有少數的測試樣本可供使用。此時如何建構在經濟且有效的逐步應力(step-stress) 或連續型應力(progressivestress)之加速衰變試驗, 是生產者所面臨的重要課題。針對非線性衰變路徑且在連續型應力加速下, 本文採用累積暴露(cumulative exposure) 模型來建構出連
續應力與正常應力之間的時間轉換函數, 進而可推估產品在正常應力下之壽命分配, 此成果對縮短衰變試驗的時間將有顯著的貢獻。

關鍵字

衰變模型 ； 產品平均壽命 ； 模型誤判分析 ； 最佳實驗配置 ； 偏斜常態 ； EM 演算法 ； 累積暴露模型 ； 連續應力加速衰變試驗

並列摘要

Degradation models have been widely used to assess the lifetime information of highly reliable products. The performance of a degradation analysis strongly depends on the modeling of product’s degradation path. For designing and analyzing the degradation tests of highly reliable products, we study the following four topics in this thesis.
(i) Motivated by a real data set, we propose a general linear degradation model in which the unit-to-unit variation and time-dependent structure are simultaneously
considered. For this model, the product’s mean-time-to-failure (MTTF) can be obtained under some regular conditions. Furthermore, we also address the effects of model mis-specification on the prediction of product’s MTTF. It shows that the effect of model mis-specification on
product’s MTTF predictions is not critical when the sample size is large enough. However, when the sample size and termination time are not large enough, a simulation study shows that these effects are not negligible.
(ii) Under the proposed linear degradation model, we study the problem of optimal test plans. Under the constraint that the total experimental cost does not exceed a pre-determined budget, the optimal decision variables
such as sample size, sample frequency and terminational time are solved by minimizing the variance of the estimated MTTF of the lifetime distribution of the product. Moreover, we also assess the robustness of this degradation model through sensitivity analysis and address the effects of variety of parameters and cost conditions on the optimal test plans.
(iii) Motivated by a laser data, we relax the normal assumption of random-effect to fit realistic data sets. In this topic, we construct a skew-Wiener linear degradation model and derive the closed-form expression of the lifetime
distribution. Because the likelihood functions for such a degradation model are analytically intractable, we develop an EM type algorithm to efficiently obtain the maximum likelihood estimators for this model.
(iv) For highly reliable products with very few test units on hand, we use the concept of cumulative exposure model to formulate a typical progressive stress accelerated degradation test (PSADT) problem. An analytical expression
of the product’s lifetime distribution can then be obtained by using the first passage time of its degradation path. Next, an analytical performance comparison between the PSADT and the constant stress degradation test
under same special cases is present. The comparison includes the product’s MTTF, median lifetime, and variations of lifetime during different stresses.

並列關鍵字

Degradation model ； mean-time-to-failure ； mis-specification analysis ； optimal test plan ； skew-normal ； EM algorithm ； cumulative exposure model ； progressive stress accelerated degradation test

參考文獻

[58] Tseng, S. T. and Liao, C. M. (1998), “Optimal design for a degradation

[60] Tseng, S. T. and Peng, C. Y. (2007), “Stochastic diffusion modeling of degradation

[1] Arellano-Valle R. B., Bolfarine, H. and Lachos, V. H. (2005), “Skew-normal

linear mixed models,” Journal of Data Science, 3, 415-438.

[4] Azzalini, A. (2005), “The skew-normal distribution and related multivariate