THE STATUS AND IMPACT OF RELIABILITY METHODOLOGY BY GERALD J. LIEBERMAN TECHNICAL REPORT NO. 115 January 2, 1969 SUPPORTED BY THE ARMY, NAVY AND AIR FORCE UNDER CONTRACT Nonr-225(53)(NR-04 2-002) WITH THE OFFICE OF NAVAL RESEARCH DEPARTMENT OF OPERATIONS RESEARCH AND: DEPARTMENT OF STATISTICS STANFORD UNIVERSITY. STANFORD, CALIFORNIA @ERRATA The attached page 35 was omitted from Technical Report No. 115 - The Status and Impact of Reliability Methodology - by Gerald J. Lieberman.(9) fo} 2) ne) Sarkar, T., Unpublished Manuscript on Reliability, Stanzord University (1968). Shooman, M. L., Probabilistic Reliability - An Engineering Approach, McGraw Hill, New York (1968). U.S. Government, "Relisbility Stress and Failure Rate Date", MIL-HD BK-217A, Government Printing Office, Washington, DC. (1965)+ Woods, W. Me, and Borsting, J. R., "A Method for Computing Lover Confidence Limits on System Reliability Using Component Failure Data with Unequal Sizes", Tethnicel Report NPS5S Wo/Bg 8061 A, U.S. Navel Postgraduate School (1968). 35THE STATUS AND IMPACT OF RELIABILITY METHODOLOGY by Gerald J. Lieberman ‘TECHNICAL REPORT NO. 115 January 2, 1969 Supported by the Army, Navy and Air Feree under Contract Nonr-225(53) (NR-O42-co2) with the Office of Naval Research Gerald J. Lieberman, Project Director Reproduction in Whole or in Part is Permitted for any Purpose of the United States Goverment DEPARTMENT OF OPERATIONS RESEARCH and DEPARTMENT OF STATISTICS STANFORD UNIVERSITY STANFORD, CALIFORNIATHE STATUS AND IMPACT OF RELIABILITY METHODOLOGY” By Gerald J. Ldeberman Stanford University System Configuration and Reliability There exist many definitions of reliability, depending upon the point of view of the user. However, they all have a common core which contains the following: Reliebility, R(t), is the probability that a device performs adequately over the interval [0,t]. Tn general, 3% is assumed thet, unless repair or replacement occurs, adequate perfor- mance at time + implies adequate performance during the interval [0,t]. Although this definition is simple, the systems to which it 1s applied are generally very complex. In principle, it is possible to break down ‘the system into black boxes, each black box being in one of two states, good and bad. Mathematical models of the system can then be abstracted from the physical processes and the theory of combinatorial probability utilized to predict the reliability of the system. Whe black boxes may ‘be very dependent on each other. For any reasonable system, such a probability analysis generally becomes so cumbersome that it mist be considered impractical. Hence, other methods ere sought which either simplify the calculations or provide bounds on the reliability of the entire complex system. Presented at the Research and Technology Conference on Optimization/ Statistics, Similation, Information Processing, eyonsored by the Office of Navel Research and the George Washington University on September 11-13, 1968.1-1 Series Systems In dividing the system into black boxes, it is often possible to structure it into a series configuretion, which is the simplest and most common of all configurations. For a series structure, the system fails if any component of the system fails. A typical example of = series structure is a string of Christmas tree Aghts clrea 1910. Suppose there exists a system composed of n components connected in series. Let X, be a binary random variable corresponding to the a°” component, and let the probability that unit i 4s successful be denoted by p, = P(X, = 1)+ Let R denote the probability that the system Will operate. Since a series system requires that each component perform successfully in order for the system to operate properly, the probability of success, e.g, the reliability, is given by R= P(X =, Using the usuel terms for conditions) probabil: ty R= PEK A2) PER TR =2) POG = HL, Heb} oes POG AL Rg) + In general, such conditional probabils fies require careful analysis. For example, P{X,-1]Xj-1} 1s the probability thet component 2 will perform successfully given that component 1 performs successfully. Consider a systen where the heat from components 1 afects the tempera- ture of component 2, and thereby its probability of success. ‘The performance of these components are then dependent and the evaluation of ‘the conditional probability is extremely difficult. yon the obher 2hand, the performance characteristics of these components do not interact, e-g-, temperature of the component does not affect the performance of the other components, then the components can be said to be independent. The expression for the reliability then simplifies and becomes R= P(X=2} P{XeL} +++ P(X 1} = pypy + Such probabilities are often readily obtainable. 1.2 Parallel systems A parallel system is defined to be a system consisting of n components such that failure occurs if all components fail. Alternatively, ‘the system operates if at least one of the n components performs satisfactorily. This property of parallel systems is often called "redundancy", e., there are alternative components, existing within ‘the system, to help the system operate successfully in case of a failure of one or more components. If the probability of failure of the i” component is denoted by P{X,=0} =1-p, then Ro LeP(K)=0, X90)0045%,20) P{X,=0} P(%=0]%4=0) P(X,=0|X,=0,Xp=0}- + +P(X,=0 |X, KyAOyseeg KO} If the unit failures are independent, then this simplifies into R= (1+p,)(1-pg) +++ (1-p,) +As an example of how a parallel structure may improve the reliability of a system consider the following exemple given by Shoonan [10], Shooman describes the reliability of a gyro system. Fe indicates that a high type militery system would use three single degree of freedom gyros, with each gyro being used to establish one of the three inertial reference axes, x, y, 2+ He assumes that each gyro costa $15,000 each 50 that the total cost is $15,000. This can be viewed as @ series system since the gyro system will work only if each of the three reference axes operate properly. Furthermore, the performance of the three gyros can be assumed to be independent, and each is known to have reliability +998. Hence the system reliability is given by 99h. P(X =1) PEK R21) POR2) = (.996)? Shoonan then points out thet $45,000 is mich too expensive for a comercial eystem. Instead, comercial systems vould use three cheaper two-degree of freedom gyros at a cost of $2,000 per unit, yielding a total cost of $6,000. Furthermore, since each gyro establishes two reference directions, there are then tvo measurenents for each reference axis. Again, the performance of each gyro reference exis is assumed to be independent but with reliability .990. Thus, the system reliability can be obtained as follows. The probability that an axis, say 2, operates satisfactorily is given by 1 - (.01)° = .9999. Hence, the probability that ell three axes operate is R= (.9999) = .9997 ,which is superior to the reliability for the military system. ‘Thus, the redundent cheaper system spears to yield a more reliable system, which is the best of all possible worlds. However, Shoomen states that "Tt is important to note that some of the cheaper gyros will be offset by nore frequent replacement and higher maintenance costs when the system is operated over a long period of time". This last statement appears to vitate the conclusions of the entire example, but it still illustrates how redundancy can improve reliability. 1.3 Standby (parallel) Model In the parallel system just described all the components are turned on at the beginning of operation, and all unite contime to perforn until they feil. When a unit fails it remains in the system which continues to operate until all units fail, or the system completes its mission. In a standby system the components are connected in parallel but do not operate at the same time. The operstion of a standby system with (n-1) spares can be described as follows: Assume that there exists a switch which can place any unit into cperation in the system. At the start of operation this switch connects to component 1 and turns ‘this unit on, Meanwhile components 2 thru n are left in reserve (standby) in @ turned-off condition. It is further assumed that the switch can detect when a component faile. When it senses improper operation, it switches to component 2, and then turns the unit on with- out interrupting the system performance. This continues through the n'” unit. The n channel standby system can fail only if all channels have failed after being activated. Hence,R= L-P(%=0, %560)+6)%,-0} = L-PEK,=0} PEK, -O[x,-0) FEO |x4-O, x0} ++ PX, 2019520) 90, 25, 10) Although this expression appears to look Like thet for the regular parallel case, its interpretation is quite different eince only one component is operating at a time. Tims, unit j begins operation only efter unite 1,2,-..,J-1 have failed. Hence, even when the performance characteristics of each unit do nct interact so thet they can be con- sidered to be independent, an elternative expression for reliability other than that previously mentioned, is required for reliability eval- uation purposes. For such independent units, an expression for the reliability at time t for a standby operation is given by R(t) = Pt, +2, + +++ + D> +) where T, 4s the time to failure of unit 1. 1.4 r Out of n Configuration Some systems have a configuration such that the system operates if r out of n components function properly. A simple example is the cables thet support s suspension bridge, 1.e., the bridge will remain erect if r cables will support the bridge. If each component is ddentical, the reliability is given by the binomial expression R= S Qstaer* , &where p ie the probability of success of an individual component. If ‘the components differ, the expression for the relisbility is more complex but can be obtained by brute force (direct enumeration). 1.5 Bounds for Reliability of Series and Parallel Systems It is evident thet for systems which are in series or parallel, the expressions for the system reliability is simplified for systems which are composed of independent components. For systems vhich are dependent, evaluation of the reliability depends heavily on knowing conditional probabilities which are very complex. Hence, obteining bounds for dependent systens based upon knowing the probsbilities of individual components is worthwhile. In order to do this it is helpful to define the concept of associsted binery random variables. This definition was introduced by Esary, Proschan and Walkup [5], and is as follows: are associated binary random variables if, and ys Lygevesl, only if covtr(r), A(t)] 2 0 for ell pairs T, A of binary, non decreasing functions. This is the generalization of the usual definition for two random variables being associated. For example, if X and Y are tyo binery randon variables Cov[X,¥] > 0 implies that X,Y are associated. The bounds can nov ve given as follows. Hf %) hy ave associated binary random variables then R(t) = POHL, Xely-+-,X a1} > i P(xj=1} = =.This states that for a binary series system in which Xj,-.-, associated, the system reliability is at least as large as the product of the component reliabilities. Another bound follows from the result which states that if %y Xyp-e+yX, ave associated binary random variebles, then 2 pK 0} > 1 PEX,=0} This states that for a binary parallel system in which Xj, Ky-+-)%, are associated, the system reliability is no greater than one minus the product of the component failure probabilities. Bounds for complex structures can also be obtained for associated random variables by approxinating these structures with a structure having only series and parallel arrangements. Such results are also given in [5]- The results of this sub-section depend heavily on the condition that the sequence of random variables are associated. Unfortunately, ‘there is no vorking method for ascertaining whether or not euch a system has associated components. For two components association is equivalent to having a non negative covariance. However, for three or more components there is no naturel extension that enables one to make rationel assumptions sbout this property for a system. 1.6 Practical Considerations Affecting Measurement of System Reliability ‘The previous sections considered reliability fron the design point of view, and indicated that, for very complex systems, the measurement of system reliability as a function of component reliabilities is very 8cumbersome and often unpractical. Bounds on the system reliability, however, can be obtained based upon a knowledge of the component re- iebilities. ven such a scheme requires information about component, reliabilities, and the question arises as to hoy to obtain such data. Unfortunately, these data are often taken from such documents as MIL-STD 217 [21] and the "Failure Rate Data Handbook" (PARADA). These data tend to report failure rates, and are usually compiled under a wide variety of conditions. To compensate for the latter, efforts are made to bolster the theory by incorporating stress factors and environmental factors. Unfortunately, the method of choosing these factors, and the subsequent use of results based upon these factors, is tantamount to using witch craft. Finally, as will be noted subsequently, reporting of the failure rate, or alternatively the mean time to failure, may not be sufficient to describe the data for reliability purposes. 2. Characterization of Reliabilities ‘The previous material has described system reliability in terms of component reliability for important classes of systems. Existence of probabilities of success hes been assumed but very little has been said about estimating their values. In fect, in both government and industry contracts are often called for which require reliability at a fixed level, say «95, be guaranteed for systems where only one of a kind will be constructed. Hence, no operational testing can be made since in many cases testing is destructive. Hence, reliability has to be estimated based upon previous experience with similar type components. In such situations engineers may be reluctant to state reliabilities for componentsin terms of probabilities but are often willing to give estimates for mean ‘time to failure. Hence, bounds on system reliability based upon this type of component data is useful, and the following material is motivatea by this point of view. 2.1 Monotone Failure Rates An appealing intuitive property in reliability is the failure rate function. Consider a component (or unit) and its associated random variable, 1, time to failure. Denote the failure distribution by F and the density function by. £. In terms of the previous discussion, define the random variable X= x(t) which tekes on the values ° ae Dect Then R(t) = Ppa) =p = a(t) [ear The failure rate function r(t) is defined for those values of t for which F(t) <1 by x(t) = 2(t)/R(t) ‘This function has e useful probabilistic interpretation; namely, r(+)at represents the conditional probability that an object surviving to age % will fail in the interval [t,ttdt]. This function is sometimes called the hazard rate. 20In many applications there is every reason to believe that the hazard function tends to inerease because of the inevitable deterioration which occurs, Such @ hazera function which rewains constant or increases with age is said to have en increasing failure rate (FR). In some applications the hszara function tends to decrease. It would be expected to decrease initially, for instance, for materials that exhibit the phenomenon of “work hardening". Certain solid state electronic devices are also believed to have a decreasing failure rate. Thus, a hazard function which remains constant or decreases with age is said to have a decreasing failure rate (IFR). 2.2 Bounds on Relisbility Under either IFR or DFR assumptions, it is possible to obtain sharp bounds on the reliability in terms of mozents and percentiles; in particular, such bounds can be derived from stetenents based upon the mean time to failure. The feilure rate function possesses some interesting properties. ‘The failure rate distribution is completely determined by the failure rate function, In particular 4t 4 easily show that, * le) = 2-R(2) = e[-[ son : Furthermore, @ distribution ie IFR (FR) if + -1nk( +) f r(E)de ais convex (concave). As en example, consider @ comporent hose feilure @istribution is given by the exponential, i-e., +/0 F(t) = P(x en*/? >0 t>e and the inequality is sharp i.e., the exponential distribution with mean @ attains the lower bound for t <0, and the degenerate distri- ‘bution concentrating at 9 attains the lover bound for t > 9. This can be represented graphically in Figure 2.1. 12t Fig. 2.1. A Lower Bound on Reliability for IFR Distributions The previous results could also provide lover bound for R(t) whenever t> 0 provided that R(t) ana 0/9 crocsed at the point + = 6. Unfortunately this is not true and all that can be said is that the cross- over is to the right of t An upper bound [2] can be obtained from the following result. If F is IFRwith mean 9, then R(t) <2 for + <0 se sor t>0 where w depends on t and satisfies 1 - 09 = can be bounded as shown in Figure 2.2 Fig. 2.2. Upper and Lower Bounds on Reliability for IFR Distributions B2.3 Increasing Failure Rete Average Now thet bounds on the reliebility of @ component have been obtained, what can be said about the preservation of monotone failure rate, i.e., what structures have the IFR property when their individual components have this property? Series structures of independent IFR (DFR) components are also TFR (FR). Standby systens consisting of a single IFR unit supported by (n-1) spaces is also IFR. k out of n structures consisting of n identics] independent components each having an IFR failure distribution is also IFR. ‘However, parallel structures of independent IFR components are not IFR unless they are identical con ponents. Standby structures of DFR components are not necessarily DFR. Thus, we see that for somewhat simple systems there my not be a preservation of the monotone failure rate. Instead of using the hazard rate as a means for characterizing the reliability t R(t) = enff soa] ’ a somewhat less appealing characterization can be obtained from the failure rate average function, * fo sae slog R(t) t = . A @istribution ¥ such that ¥(0) = 0. is called IPRA (increasing * failure rate average) if end only is[ f 2(s)ag]/t 4s non decreasing © in t>0. A similar definition is given for DFRA. Birnbeum, Beary ayand Marshall [3] showed that the IFRA class of distributions is closed under the formation of monotone binary systems. Thus, complex systems composed of components which are exponential (which is IFRA since IFR Gistributions are also IFRA) are also IFRA. As with IPR systems there ere bounds for systems vhich are IFRA. In fact, the same upper bound as given for IFR distributions is still applicable here, Barlow and Marshall [1] give the lower bound for IFRA distributions with mean 6 as follows: R(t) > 0 if tpe eb gt/9y R(t) > minfe for t0, where 8 is called the shape parameter and @ is called the scale parameter, Note that when 6 =1, the Weibull reduces to the exponential. The C.D.F. for the Weibull is given by re) #2 = (H/o? ‘ Vv ° =o t1, x(t) isa monotonically increasing function of t, and for 6 <1 is @ monotonically decreasing function of t. This is shown in Figure 3.1. 20ae Fig. 3.1. Hazard Functions for the Weibull Distribution There is considerable statistical literature on estimating reliability based on the Weibull distribution. However, until @ recent paper by Johns and Lieberman [6] there has been no satisfactory method presented for computing lower confidence bounds for reliability which are exact for each sample size n when both @ and f are unknown parameters. In fact, prior to the aforementioned paper, mich of the Literature concerned itself with finding lover confidence bounds when ‘the shape parameter 8 was assumed to be known, @ very unrealistic assumption. The Department of Defense has issued three pamphlets which are based upon this assumption of known shape parameter. It is cesily shown that if one assumes that § is known, this problem reduces to the case of the exponential distribution. This is easily seen by operating on the traneformea variable cy which has en exponential distribution with pareseter @ = OP. Hence, alif X has a Weibull distribution with parameters a@ and B (known) 6 R(t)=e % =Pz>thy, where 2 =x? has an exponential distribution with psremeter 9 = a. ALL the deficiencies of the exponential assumption are really present, although somewhat obscured, when the shape parameter ® 1s ascumed known, When both @ and B are assumed unknown, the traditional methods for computing lower confidence bounds lesd to analytical difficulties which preclude getting exact solutions. The Johns-Lieberman paper gives the first satisfactory method for computing exact confidence bounds. This paper presents a simple method for obtaining exact lover confidence pounds for reliabilities for items whose life times follow a Weibull Gistribution where both the scale parameter and shape parameter are unknown, These confidence bounds are obtained when the data has deen subjected to Type II censoring at r out of n, and are asymptotically efficient. They are exact even for small ssmple sizes in that they attain the desired confidence level precisely. ‘Tables are given of exact lover confidence bounds for the reliability for sample size of 10, 15, 20, 30, 50, and 100 and for verious values of r (including r =n) and for various confidence coefficients 7. Because of the difficulty in obtaining these bounds analytically, & simlation wae required for the construction of these tables. Asymptotic expressions for the lover confidence bounds which are analytically tractable are also given. 223.3. General Failure Rate Functions The Weibull distribution is considered to be an important time to feilure distribution since it is a "rich class" i.e., it is a two parencter family which seems to fit many sets of deta. In fact, on the basis of sample data it would be virtually impossible to conclude that the tine to failure distribution aid not cone from a Weibull distribution. However, there are other families of distributions with simlar properties such as the Gama, Log Normal, ete. Another method for describing time to failure distributions is to investigate the failure rate function. It has been pointed out that ‘ R(t) = 1 - F(t) = exp Sf 2()8f yor alternatively , ° f r(g)ag . F(t) <1 - xp The simplest form of function for r is to assime that r is constant. This is the case of the exponential previously zentioned. When wear of deterioration is present, the hazard function will increase as tine passes. The simplest increasing-hazard wodel that can be postulated is the linear function r(t) = Kt for +> 0. This leads to a time to failure distribution of the form 2 w(t) = 2 - oR? which is the form of the Raleigh distribution. Another type of failure rate function to consider is given by r(t) = Kt" which is equivalent to the Weibull distribution. 23In sone cases the failure rate function is initially constant and then begins to increase rapidly. This could be represented by a combination of a constant and a linearly increasing hazard or an expo- nentially growing hazard. The hazard function r(t) = Ke leads to the extreme value distribution for the time to failure, i-e., F(t) = 1 - expl-(K/a)(e-2)1 - Depending upon the choice of parameters, the failure rate functions usually lead to monotone functions. However, in many practical appli- cations, the so called bath-tub function is desirable. Such a failure rate function appears in Figure 3.2. r(t) Fig. 3.2. The Bath Tub Hazard Function A hazard function which leads to this bath tub shape is given by 2h(4) = pf (at)? o and the corresponding CDF is given by =1-2 we Mau F(t) = 1 = sty 5 eau. at This is called a mixed Weibull-Gamma distribution. When y =1, the Weibull is obtained. When 6 =1, the Gamma is obtained. 3.4 Practical Considerations Related to the Statistical Estimation of Reliability There exists statistical procedures for obtaining lover confidence Limits on component reliability assuming a paranetric form for the under- lying time to feilure distribution. This requires obtaining data from ‘tests, many of which are performed in a laboratory or simulated environment. Hence, these lower confidence limits are often subject to question. In addition, the assumption of a paremetric form for the underlying time to failure distribution requires serious study. All too often, this assumption is made without due consideration of its implication. Assuming an exponential implies no deterioration; assuming a Weibull implies a monotone hazard function, etc. In general, the Weibull distribution is a "rich" class of distributions having a monotone hazard function, i.e., it is a two parameter family having great flexiility. It is almost impossible to view some sample data and conclude that it does not come from the class of Weibull distributions. (which includes the exponential 25@s a special case). Of course, other two parameter families have this property. Finally, lower confidence bounds for reliability are usually readily obteineble for components. However, even for simple systems where the form of the time to failure distributions are assumed mown, there exist few results for finding lover confidence bounds for the overall system based upon component data. 4, Estimation of Reliability Based Upon Attributes 4,1 Lower Confidence Limits ‘The previous material on statistical estimation of reliability was devoted to a discussion of time to failure data. If this type of data is not available, and instead binary data (success or failure) is recorded, statistical estimates of reliability can still be obtained. Let X be @ binary random variable. If the system (or component) performs satisfactorily during e test, the random variable X takes on the value 1. If a failure occurs, the rendom variable X takes on the value 0, Assume that P(X=l}=p and P[x=0}=1-p s0 that the relisbility is just thé probability of @ successful test, Le, ‘Thus, estimating R is equivalent to estimating the parameter p of a Bernouilli random variable. Suppose a sequence of n independent Bernoulli trials with parameter p is made and the random variables 26Xs XypreeyH, axe observed. It is well known that, EX,, which 7 corresponds to the total number of successes, tas a binomial distribution with parameters n end p. The value of p hich satisfies the expression isa lower 7 confidence bound for -p. Suppose a series system consisting of k independent components is of interest. Further, tests are to be made on each of the k com- ponents, and a confidence interval estimate for the system reliability is desired. This ds still an open research problem. Several approximate solutions exist, and one that is fairly good ie as follows. Assume that the reliability for each component p, is fairly close to 1, and the number of tests on each component n, is fairly large and equal to a , ee ee eee : reliability where 1,2 aa(D) = 3 4, so(pe1) and Cyjotoma) 18 Me Ley Percentile of the eh ayunze distribution with 2(DHl) degrees of freedom and D represents the total number of teiiures obeerved, totaled over ell Kk comporents. An alternative conservative approximation is given as follows Denote ty Py, Pys+-+yP, the PV® rower confidence bound for eTeach component respectively (found using the binomial expression given reliability of the ayaten. Ina recent paper, Borsting and Woods [12] present an approximate nethod for constructing system reliability using comporent failure date vhen the sample sizes for testing on the component parts differ greatly. This is still an area where further research may prove fruitful. 4,2 Practical Considerations Related to Obtaining Lover Confidence Limits Using Attributes Data In performing tests on complex systems, these systems are tested, modified, redesigned, etc. depending upon the outcomes of the tests so ‘thet true rendom sample is never obtained, thereby ovening the theoretical results to question. In actuality, there is a reliability growth taking place which is often ignored in reporting lower confidence dounds. Relisbility growth generally occurs on all programs, and should ‘be included in reliability estimation. Underlying Physical Models of Random Variables Section 5 was concerned with estimating the reliability of a component or a system given the parametric form of the tine to failure distribution. It was pointed out that an exponential vime to failure distribution leads to a constant hazard function, which in turn implies that "no aging" takes place. However, this distribution can be obtained by making other (equivalent) assumptions about the underlying physical process. The purpose of this section is to look at such assumptions for a variety of processes. The alternative to this approach is to look at 28sample data, and thereby assume the parametric form of the distribution of the random variable; often an unsatisfactory procedure. 5.1 Exponential Failure Law Suppose that a component fails because of the appearance of un- predictable random disturbances. These may be caused by the appearance of certain external forces Like sudden changes in the environment, or the appearance of certain intemal forces such as a malfunctioning part. Denote ty x(t) the mmber of such @isturbances occurring up to time ‘. Further, assume that for any fixed t, the random variable x(t) has a Poisson distribution with paraucter t/ (Poisson Process). Suppose that the component faile during the time period [0,%] if at last one such disturbance occurs. Under these assumptions, it ie easily shown that the distribution of the time to failure is exponential. 5-2 Gamma Failure Law Suppose the model of Section 5.1 is generalized as follows: Ac before, assume that the disturbances occur according to a Poisson process. However, now assume that the component fails whenever k or more dis- turbances occur during the time interval ([0,t]. Thus, the CDF for the time to failure distribution is given by F(t) = 1 - P(t >t} Since T>+t if and only if (k-1) or fewer disturbances has occurred, it follows that 29ker (2/oy—"t/? F(t) <2 - i > fs which is the CDF of the Gamma distribution. 5-3 Weibull Feilure Law Assume that a component (or system) can be characterized by a chain made up of identical links in the sense that the distributions of breaking strengths of each link in the chain are the same. The chain breaks (component fails) when its weakest link fails. A stress is applied ‘to the chain as @ whole and it is assumed to be applied equally to each link. Thus, the probability distribution of the time to failure of such a component ie just the probability distribution of the first order statistic, the minimm. If F represents the CDF for each link, and G represents the CDF of the chain consisting of n links then n G(t) = 1 - [2-F(t)]" . Tt is easily seen that if F is Weibull then so is G. In fect, if F(t) ~ ct for t close to zero (@>0) then G has a limiting Weibull distribution. This includes the exponential, Weibull, Beta etc. 5.4 Log Normal Failure Distributions A model which leads to the Log Normal failure distribution will be described in terms of a fatigue crack. Let X s, Y>t} = P[U>s, V>t, W> mx(s,t)} -s/0, -t/6-nax(s,t)/6. = P{U> 5} P[v >t} P{W > max s,t} =e 5 The reliability at time % 4s then easily found, end given by ~(1/0,41/0,41/8,,)t R(t) = P[X>t, Y>+] Bee ae A which is called the bivariate exponential distribution. Marshall and Olkin have also considered a mitivariate exponential distribution which is essentially @ generalization of the above arguments. 5.6 Practical Considerations Pertaining to Physical Nodels As indicated earlier, the arbitrary choice of a yarametric form of the underlying time to failure distribution is often an unsatisfactory procedure. An alternative to this is to investigate the underlying physical process which induces the time to failure distribution. This 32may lead to en appropriate choice of the parametric form of this distribution. Of course, no parametric form should be assumed, no matter hov reasonable the derivation from physical law appears, until it is compared with actual data. 33Q) [2] 3] tf] [6] (7 (8) BIBLIOGRAPHY Barlow, R. E., and Marshall, A. W., "Bounds on Interval Probabilities for Restricted Families of Distributions", 5‘ perkeley Symposium (1965) pp. 1229-257. Barlow, R. E., and Prochan, F., Mathewatical Theory of Reliability, John Wiley and Sons, New York, 1965. Birnbaum, Z. W., Esary, J. D-, and Marshall, A. i, "A Stochastic Characterization of Wear-Out For Components and Systeme", Annals of Mathematical Statistics, Vol. 37, No. 4 (1966). Birnbaum, Z. W., and Saunders, S. C., "A New Family of the Distributions", Technical Report No. 52, Leboratory of Statistical Research, University of Washington (1968). Beary, J. D-, Prochan, F., end Walkup, D., "A Multivariate Notion of Association, with Applications", Annals of Mathematical. Statistics, Vol. 38, (1967) pp. 1s66-2474. Johns, M. V., and Lieberman, G. J., "An Exact Asymptotically Efficient Confidence Bound for Reliability in the Case of the Weibull Distribution", Pechnonetries, Vol. 3, No. 1 (1966) pp. 155-175. Kramer, H. C., "One-Sided Confidence Interval for the Quality Indices of @ Complex Item", Technonetrics, Vol. 5, (1965) pp. 400-103. Marshall, A. W., and Olkin, I., "A Multivariate Exponential Distribution", Journal of the American Statistical Association, Vol. 62, (1967) pp. 30-bi. Bh1 UNCLASSIFIED ) security Clase DOCUMENT CONTROL DATA- R&D (Security cauatiction of te, body of abetact and interna annotation mut be entred nian te overall opera cl Stanford University UNCLASSIFIED Dept of Operations Research & Dept. of Statistics [2 enoue Stanford, California 9305 Tae Status and Impact of Reliability Methodology } -eBESCRIPTIVE WOVE (hype ol vapor od ince aaa) ‘Technicel Report | T RERORGY tran Wan TTT Lieberman, Gerald J. 35 i 53) | : ormonees 60 Technical Report Now 135 a-0l2-c0e 2 January 1969) | Distribution of this document is unlimdi Logistics & Mathematical Statistics Branch Office of Navel Research |Weshington, D.C. 20560 This paper presents the current status of reliability methodology, and describes its practical impact. Nathematical models of complex systems are analyzed, together with the conbinational probability necessary to calculate the system reliability, or bounds on the system reliability. Component time to failure distributions are studied and lower confidence bounds obtained Models of underlying physical processes which induce time to fallare distributions are also Anvestigatea. DD .59%. 1473 UNCLASSIFIED Security ClassificationUNCLASSIFIED ty Classification Reliability Series System Parallel System Standby System Associated Random Variables Increasing Failure Rate (IPR) Decreasing Failure Ratt (IPR) Failure Rate Hazard Function Increasing Failure Rate Average (IPRA) Decreasing Failure Rate Average (IPRA) Exponential Weibull, Gamma Bivariate Exponential INSTRUCTIONS 1, ORIGINATING ACTIVITY: Enter the name of the contractor, subcontractor, grantee, Dep! fense activity of other ongentsation (corporate author) isauing the report 2a, REPORT SECURTY CLASSIFICATION: Enter the over 2 security claseification of the sepor. ndicate whether cStricted Data’ fe included Marking ta to be in accord lance with appropriate security regulations: 26, GROUP: Automatic downgrading is specified in DoD Di- fective $200.10 and Aresed Forces Industriel Manual. Enter {he group nunber, Also, whem applicable, show thet optional Mngt have been une Yor Group i and Group 4 65 author ort ttle In all F5 meaningful tle cannot be selected without classifice ification in all capitals ta parenthesis 4. DESCRIPTIVE NOTES If appropriate, enter the type of Spon, ego interim progress, summary, annual, of fi {Glow the fhclusive dator when a specific reporting period is 5. AUTHOR(S} Enter the name(s) of inthereport, Enis taat name, te If military, chow tank and brench of services) The the principal author w en abeclute minimus requ 6 REPORT DATE: Enter the date ofthe report Sonth, peor, ormonth, yest If more than one date appears a the report, ue date of publicetion. 7a, TOTAL NUMBER OF PAGES: The total page count Should follow nonal pagination procedures, Le. eater the umber of pages contsiming information. 7h, NUMBER OF REFERENCES: Enter the total number of cferonces cited in the reper >, fo, & 8d. PROJECT NUMBER: Enter the appropriate rilitety department identification, such ex project number, ‘Riprsject number, system aumbers, task number, ete 90. ORIGINATOR’S REPORT NUMBER(S): Enter the off: Sias feport nonber by when the document will be idetified Se SARE Ye rinatng ect. “Tae number mot ‘9b, OTHER REPORT NUMBER(S): If the report has been signed any other report mumbere (either by the originator Sry the sponsor), albo enter this momber()- 10, AVAILABILITY/LIMITATION NOTICES: Enter any lime itations on further dissemination of the report, other than those] DD 1473 (BACK) om inpgted by secuty clesifcstion sing staan statemente (1) *Quaitied rea repor from BBC “Foreign announcement and diesgalnation of this fepon by DDC ia not authorized” “U, S. Government agencies may obtain copies of this report directly ftom DDC. Other qualified DDC Shere thall requewt through lore may obtain coples of this @ o "U, & military agencies may obtain copies of this tepon directly from DDC. Other quelified users ‘hall request tiorgh Ail distribution of this report ia controlled Qual itled DBC sere shall request through ° I¢ the repent hes been farahed to the Office of Techaleal Services, Department of Commerce, for eate tothe public, indi Sate ths fact and enter the price, if knows 11, SUPPLEMENTARY NOTES: Use for additional explant 12, SPONSORING MILITARY ACTIVITY: Ener the name of the depertmentel project olfice or laboratory sponsoring (PeP™ Ing fob the research and development. Include adaress. 13. ABSTRACT: Enter an abstract giving « brief and factual ‘ummary of the document Indicative of the report, even though Imey also sppser elaewhere in the body ofthe technical ree por If additional space is required, « continuation sheet shall tis highly desirable thatthe abstract of classified reports bee unclassified. ‘Each pamereph ef the abstract shall end with th indication of the military security classification ofthe in- FSnmetion in Use paragraph, repremented as (15). (8) (0). 0¢ (W) ‘There is no limitation an the length of the abetract. How fever, ihe nuggeated length iv from 180 00 228 words. 14. KEY WORDS: Key words are technically meaningful terme r'ahor! phrases tha! characterize « report and may be used a Index entice for cataloging te report, Key words must be elected no that ne seciriy classification te requlced. Ident Tisre, such ae equipment model designation, tage name, military project code name, guographle location, may be ed au key. Srords but sil be followed by an indicetion of technical con- fonts The susignment of links, ralos, and weights te optional UNCLASSIFIED Sectrily Classification