MTTF,Failrate,Reliability,and Life Testing

1 At Burr-Brown, we characterize and qualify the reliability ofour devices through high temperature life Testing . The re-sults of this Testing are quantified with such values as MTTFand failure rate. This information can be very valuable whenused for comparative purposes or applied to reliability cal-culations. However, this information loses its worth if it isnot precisely understood and appropriately employed. It isthe intent of this application note to bring together, in aconcise format, the definitions, ideas, and justificationsbehind these reliability concepts in order to provide thebackground details necessary for full and correct utilizationof our life Testing PRELIMINARY DEFINITIONSR eliability is the probability that a part will last at least aspecified time under specified experimental conditions (1).

2 mttf is the mean time to the first failure under specifiedexperimental conditions. It is calculated by dividing the totalnumber of device hours by the number of failures. It isimportant to note, at this time, that the dimensions of MTTFare not hours per failure, but rather, device hours perfailure. If each part has a chance of failure before 1hour then 10 parts have a 1% chance experiencing a failureby that time. The mttf will be the same in both cases. 1failure in 10 hours on 1 part or 1 failure in 1 hour on 10 partsboth produce an mttf of 10 device rate is the conditional probability that a device willfail per unit of time. The conditional probability is theprobability that a device will fail during a certain intervalgiven that it survived at the start of the interval.

3 (5) Whenfailure rate is used to describe the frequency with whichfailures are expected to occur, the time units are typicallydevice is simply failure rate scaled from failures per device hour to failures per billion device TO THE DETAILSIn the definition section mttf is defined as the averagetime, in device hours, per failure observed under specificexperimental conditions such as a life test. Here at Burr-Brown we use a slightly modified formula for mttf . Wecalculate 2 times the total device hours, Tdh, divided by theupper 60% confidence limit of a chi-square distribution with2 times the observed number of failures + 2 degrees offreedom, X2(2f + 2). Our formula isSince both time and failures are doubled, these definitionsare roughly equivalent.

4 Some explanation is in multiple life tests are run on the same type of device, it isunlikely that all tests will have the same number of failuresfor the same number of device hours. Rather there will bea distribution of failures. The minimum value must be 0 forno failures. The maximum value could correspond to 100%failures, but we can presume that we are running enoughparts that this will not happen. Rather the distribution willtaper off as the number of failures increases. Somewhere inbetween there will be a concentration of chi-square calculation provides us with a tool for adjust-ing the actual number of failures from a limited life test tomake it more accurately reflect what we might expect fromthe population as a whole. For example, applying a confi-dence level of 60% to a chi-square distribution with 8degrees of freedom will return a value into the denominatorof the mttf calculation which is greater than or equal to60% of the values in a chi-square distribution with a meanof intuitive interpretation of the chi-square calculation isthat the calculated value represents, roughly, a number offailures which will be greater than 60% of the failures wemight get during multiple life tests.

5 The upper 60% level isselected because it represents an approximately averageestimate for mttf and because it is widely accepted amongsemiconductor manufacturers and users. This method ofestimating mttf does not prevent further reliability calcu-lations from being made at more conservative more point remains to be explained regarding thiscalculation. Why do we use 2 (# failures) +2? The technicalexplanation for this is given later in this paper. Briefly, thefactor of 2 is necessary to achieve theoretical validity of theX2 distribution. Given the factor of 2, it can be seen that weare merely adding 1 failure to the actual number of added failure appears in the calculation as if a failureoccurred at the end of the test. This assures that the testterminates with a failure, also a theoretical requirement, aswell as allows calculation of mttf even if no failures mttf value by itself really only serves for comparisonpurposes.

6 Many more factors need to be considered beforepredictive statements regarding the longevity of our compo-nents can be made. The statistical concepts of reliability andfailrate allow us to make such predictions. I will presenthere, with justification yet to come, the statistical formulaswhich quantify these 2(2f+2) mttf , failrate , reliability AND life TESTINGby Bob SeymourAPPLICATION BULLETIN Mailing Address: PO Box 11400 Tucson, AZ 85734 Street Address: 6730 S. Tucson Blvd. Tucson, AZ 85706 Tel: (602) 746-1111 Twx: 910-952-111 Telex: 066-6491 FAX (602) 889-1510 Immediate Product Info: (800) 548-6132 1993 Burr-Brown CorporationAB-059 Printed in December, 1993 SBFA0112 Let reliability be represented by R(t) and failure rate by Z(t).ThenandRemember, reliability is the probability that a part willfunction at least a specified time.

7 Failure rate describes thefrequency with which failures can be expected to occur. Byexamining failure rate we can make important statementsabout the life cycle of the life cycle of a part can be thought of as having threedistinct periods: infant mortality, useful life , and three periods are characterized mathematically by adecreasing failure rate, a constant failure rate, and an in-creasing failure rate. This theory is the basis of the ubiqui-tously discussed bathtub curve .The listed formulas can model all three of these phases byappropriate selection of and . affects the shape of thefailure rate and reliability distributions. When < 1 Z(t)becomes a decreasing function. = 1 provides a constantfailure rate.

8 An increasing failure rate can be modeled with > 1. Therefore, can be selected to accurately model theshape of an empirically known failure rate (or of the originalprobability density function of T which defines the failurerate). The constant provides the scaling good design, debugging, and thorough Testing ofproduct the infant mortality period of a part s life should bepast by the time the parts are shipped. This allows us to makethe assumption that most field failures occur during theuseful life phase, and result, not from a systematic defect,but rather from random causes which have a constant failurerate. The constant failure rate presumption results in = concept of a constant failure rate says that failures canbe expected to occur at equal intervals of time.

9 Under theseconditions, the mean time to the first failure, the mean timebetween failures, and the average life time are all , the failure rate in failures per device hour, is simplythe reciprocal of the number of device hours per isduring constant failure rate that mttf is always the number of device hours perfailure but neither failure rate nor is always 1 DERIVATIONS AND JUSTIFICATIONSOK, it s time for some real details. Virtually all of theinformation on the Weibull distribution comes from Prob-ability and Statistics for Engineers and Scientists by RonaldE. Walpole and Raymond H. Myers, copyright 1985,Macmillan Publishing Company. Much of the informationin the section on mttf is extrapolated from the lectures ofR(t)=e t Z(t)= t 1Z(t)= Z(t)= 1 / MTTFDr.

10 Duane Dietrich, professor of Systems and IndustrialEngineering at the University of Arizona. My apologies toDr. Dietrich for any s start by hypothetically running a huge life test longenough to drive all devices to failure, recording time-to-failure for each part, generating histograms, calculatingMTTF, etc. A histogram of time-to-failure would be shape is unknown and unimportant at this time. Given thehypothetical nature of the experiment, we can presume thatthe distribution is representative of the whole this distribution, we can describe reliability asR(t) = P (T > t)where T represents time-to-failure and t represents that this is merely an exact restatement of the verbaldefinition already useful function which can be derived from the time-to-failure histogram will represent the cumulative probabil-ity of failure at any time t.