周偉萍,王豐效
(喀什師范學(xué)院數(shù)學(xué)系,新疆喀什844006)
雙參數(shù)指數(shù)分布步加試驗(yàn)TFR模型下修正的MLE
周偉萍,王豐效
(喀什師范學(xué)院數(shù)學(xué)系,新疆喀什844006)
給出了雙參數(shù)指數(shù)分布全樣本場合下步進(jìn)應(yīng)力加速壽命試驗(yàn)TFR模型下參數(shù)的修正極大似然估計(jì),并通過Monte-Carlo模擬證明了修正的極大似然估計(jì)要好于極大似然估計(jì).
雙參數(shù)指數(shù)分布;步加試驗(yàn);TFR模型;極大似然估計(jì)
目前對步進(jìn)應(yīng)力加速壽命試驗(yàn)TFR模型的統(tǒng)計(jì)分析已有不少研究,具體可參閱文獻(xiàn)[1-14].文獻(xiàn)[14]給出了全樣本場合下雙參數(shù)指數(shù)分布步進(jìn)應(yīng)力加速壽命試驗(yàn)TFR模型下參數(shù)的極大似然估計(jì).雖然極大似然估計(jì)有良好的大樣本性質(zhì),但是文獻(xiàn)[14]中門限參數(shù)的極大似然估計(jì)往往偏大.本文就是基于這一事實(shí),利用糾偏的思想,構(gòu)造一新的統(tǒng)計(jì)量,給出參數(shù)的修正極大似然估計(jì),并且把改進(jìn)后估計(jì)量與文獻(xiàn)[14]中給出的參數(shù)的MLE作模擬比較,以期使改進(jìn)后估計(jì)量的相對誤差有明顯減小.
約定如下記號:用隨機(jī)變量Y表示在一個持續(xù)應(yīng)力下某產(chǎn)品的壽命時(shí)間,它的分布函數(shù)、密度函數(shù)、殘存函數(shù)和失效函數(shù)分別用和來表示,不同應(yīng)力場合下的各量,可通過添加一個下標(biāo)來表示.例如,在應(yīng)力下,將表示產(chǎn)品壽命時(shí)間的分布函數(shù).而步進(jìn)應(yīng)力加速壽命試驗(yàn)中將用來表示產(chǎn)品的壽命時(shí)間,當(dāng)指的是時(shí),以上所提到的每個函數(shù)都要加一個星號.
可以得到相應(yīng)的殘存函數(shù)
特別地,如果是考慮簡單步進(jìn)應(yīng)力加速壽命試驗(yàn)場合,即將n個產(chǎn)品,在應(yīng)力下試驗(yàn)做到時(shí)刻,緊接著將應(yīng)力提高到,試驗(yàn)做到時(shí)刻,此時(shí)TFR模型的失效率函數(shù)、殘存函數(shù)分別為
由于門限參數(shù)μ≤y<∞,所以用?μ=y1來估計(jì)μ時(shí)往往會偏大.關(guān)于門限參數(shù)μ的估計(jì)的修正基于如下事實(shí):
從而在式(1)~式(3)的基礎(chǔ)上,我們利用糾偏的思想構(gòu)造及因子α的新估計(jì):
例1[14]取樣本容量n=15,參數(shù)真值取為μ= 10,=1,α=3,另外取=5,通過Monte-Carlo模擬得到如下失效數(shù)據(jù):10.048 0,10.068 0,10.254 0,10.258 1,10.336 0, 10.411 3,10.455 7,10.461 8,10.465 9,10.550 9, 10.556 0,10.626 7,10.631 4,11.338 2,11.662 4.
模擬比較的結(jié)果見表1(其中MEL的結(jié)果來自于文獻(xiàn)[14]).
例2[14]取樣本容量n=20,參數(shù)真值取為μ=15=5,α=2,另外取=7,通過Monte-Carlo模擬得到如下失效數(shù)據(jù):
15.130 1 ,15.289 0,15.475 6,16.377 7, 16.588 0,17.352 7,17.449 6,17.586 8,17.612 3, 18.500 3,18.903 6,18.926 3,19.175 4,19.249 5, 19.725 1,19.872 2,20.022 2,20.997 9,23.898 7, 29.132 6.
模擬比較的結(jié)果見表2(其中MEL的結(jié)果來自于文獻(xiàn)[14]).
表1 MMLE與MLE的模擬比較結(jié)果
表2 MMLE與MLE的模擬比較結(jié)果
從表1和表2可以看出,在全樣本場合下從參數(shù)估計(jì)的相對誤差來看,修正的極大似然估計(jì)(MMLE)要優(yōu)于極大似然估計(jì)(MLE).
例3 為了有一個橫向的比較,針對小樣本、中樣本和大樣本這三種情況分別產(chǎn)生隨機(jī)數(shù).參數(shù)真值取為μ=25,θ1=5,α=10,在這樣的前提下,通過Monte-Carlo方法進(jìn)行數(shù)據(jù)模擬。當(dāng)取n=50(小樣本)時(shí),取r1=30;當(dāng)取n=200(中樣本)時(shí),取r1=100;當(dāng)取n=1 000(大樣本)時(shí),取r1=800.表3中給出了MLE(文獻(xiàn)[14]中的方法)和本文討論的MMLE(修正極大似然估計(jì))模擬結(jié)果.
表3 小樣本、中樣本和大樣本的MMLE與MLE的模擬比較結(jié)果
從表3可以看出,在小樣本、中樣本和大樣本這三種情況下,本文討論的MMLE(修正極大似然估計(jì))整體上要優(yōu)于MLE(文獻(xiàn)[14]中的方法),更適合研究者采用。
由模擬結(jié)果可知,本文給出的修正極大似然估計(jì)方法從相對誤差的角度來看要優(yōu)于極大似然估計(jì),因此本文提出的修正極大似然估計(jì)方法是可行的,也是比較理想的.
[1]Bhattacharyya G K,Soejoeti Z.A tampered failure rate model for step-stress accelerated life test[J].Communication in Statistics Theory and Methods,1993,22(9):2 631-2 639.
[2]Madi M T.Multiple step-stress accelerated life test:the tampered failure rate model[J].Communication in Statistics Theory and Methods,1993,22(9):295-306.
[3]Seiji Nabeya.Coincidence of two failure rate models[J].Communication in Statistics Theory and Methods,1993,22(3):781-785.
[4]Rao B,Ra J.Equivalence of the tampered random variable and the tampered failure rate models in accelerated life testing for a class of life distributions having the“setting the clock back to zero property”[J].Communication in Statistics Theory andMethods,1992,21(3):647-664.
[5]陳迪.步進(jìn)應(yīng)力加速壽命試驗(yàn)的比例失效率模型[C]//首屆中國運(yùn)籌學(xué)會青年可靠性學(xué)術(shù)會議論文集——可靠性理論、方法及應(yīng)用.北京:機(jī)械工業(yè)出版社,1994.62-66.
[6]Wang R H,Fei H L.Uniqueness of the maximum likelihood estimate on the Weibull distribution tampered failure rate model[J].Communication in Statistics Theory and Methods,2003,32 (12):2 321-2 338.
[7]Wang R H,Fei H L.Conditions for the coincidence of the TFR,TRV and CE models[J].Statistical Papers,2004,45:393-412.
[8]王蓉華,費(fèi)鶴良.TFR、TRV和CE模型序加試驗(yàn)下Weibull分布產(chǎn)品的失效分布[J].運(yùn)籌與管理,2002,11(5):47-55.
[9]Wang R H,Fei H L.Statistical inference of weibull distribution for tampered failure rate model in progressive stress accelerated life testing[J].Journal of Systerms Science and Complexity, 2004,17(2):237-243.
[10]王蓉華,費(fèi)鶴良.TFR模型序加試驗(yàn)下Weibul1分布產(chǎn)品壽命的統(tǒng)計(jì)分析[J].運(yùn)籌與管理,2004,13(2):39-44.
[11]Xu X L,Fei H L.Approximate maximum likelihood estimate and inverse moment estimates of the parameters of the tampered failure rate model for the weibull distribution in a step-stress accelerated life test[J].數(shù)學(xué)研究,2003,36(4),351-367.
[12]Xu X L,Fei H L.Parameter estimation of the weibull distribution tampered failure rate model under a normal stress[J].應(yīng)用概率統(tǒng)計(jì),2004,20(2):126-132.
[13]王蓉華,徐曉嶺,施宏偉.Gompertz分布TFR模型多步步進(jìn)應(yīng)力加速壽命試驗(yàn)的統(tǒng)計(jì)分析[J].應(yīng)用概率統(tǒng)計(jì),2009,25(1):47-59.
[14]王蓉華,徐曉嶺,劉文華,等.兩參數(shù)指數(shù)分布產(chǎn)品全樣本場合下步進(jìn)應(yīng)力加速壽命試驗(yàn)損傷失效率模型下的統(tǒng)計(jì)分析[J].運(yùn)籌與管理,2007,16(3):74-77.
(編輯:郝秀清)
The modified MLE of parameters of two-parameter exponential distribution based on TFR model under step-stress accelerated life testing
ZHOU Wei-ping,WANG Feng-xiao
(Department of Mathematics,Kashi Teachers College,Kashi 844006,China)
We obtain the modified maximum likelihood estimator of parameters of two-parameter exponential distribution with full sample size based on tampered failure rate model under stepstress accelerated life testing,and show that the modified maximum likelihood estimation is better than the maximum likelihood estimation through the Monte-Carlo simulation.
two-parameter exponential distribution;step-stress accelerated life testing;tampered failure rate model;maximum likelihood estimator
1672―6197(2013)01―0030―04
O213.2
A
2012- 12- 26
喀什師范學(xué)院一般課題((11)2389)
周偉萍,女,470988895@qq.com