KEMBAR78
Practical Reliability Engineering | PDF
Scribd
Upload
2K views246 pages

Practical Reliability Engineering

Confiabilidade

Uploaded by

Sandro Souza
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
2K views246 pages

Practical Reliability Engineering

Confiabilidade

Uploaded by

Sandro Souza
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
You are on page 1/ 246
1 Introduction to Reliability Engineering WHAT IS RELIABILITY ENGINEERING? [No one disputes the need for articles tobe relable. The average consumer is acutely Aware of the problem of ess than periect reliability in domestic products such as TV sets and automobiles, Organizations such as airline, the military and public utilities are aware of the cosis of unreliability. Manuiacturers often suffer high costs of failure under warranty. Argument and misunderstanding begin when we try to quantify celability values, or try to assign financial or other benefit values to levels, of reliability “The simples, purely produceroviented or inspectors’ view of reliability i that in which a product is assessed against a specification or set of attributes, and when pasted is delivered to the customer. The eustomer, having accepted the product, ‘Sccapts that it might fail at some Future time. This simple approach is often coupled ‘with a warranty, or the customer may have «ome protection in law, 9 that he may ) e— Mesenat erase caste ragaronens 7, Lega eutuony Put att Deveepren ie Agere 11 reception of rk Later chapters wil show how reliability engineering methods can be applied to design development and management to control the level of risk The extent to which the methods are applicable must be decided for each project and for each design area. They must not replace normal good practice, such as safe design for components subject to cyclic loading, or application guidelines for electronic ‘components. They should be used to supplement good practice. However, there ‘are times when new risks are being taken, and the normal rules and guidelines are inadequate or donot apply. Sometimes we take risks unwittingly, when weassume thatwe can extrapolate safely from our present knowledge Designers and managers sre often overoptioiaic ar are rictant ta point cut risks about which they tre tis for these reasons that an understanding of reliability engineering principles ‘and method's i now an easential ingredient of modern engineering. WHY DO ENGINEERING ITEMS FAIL? ‘There are many reasons why a product might fail. Knowing, os fra is practical the tential causes of failures is fundamental to preventing them. Its rarely practicable to anticipate all of the causes, so its also necessary to take account of the uncertainty involved. The reliability engineering effort, during design, development and in ‘manufacture and service should address all ofthe anticipated and possibly unantici- [pated causes of failure, to ensure that their eceurrence is preventad or minimized. ‘The main reasons why failures occur are: 1, Thedesign might be iahrestly incapable. might be oo weak, consume too much power, siffer resonance atthe wrong frequency, ec. The list of possible reasons is tendiess, and every design problem presents the potential for err, omissions, fand oversights. The more complex the design cr dificalt the problems to be ‘overcome, the greater is this potential. Why do Engnmeing Ime Fail? 5 2. The item might be overstressed in some way. If the stress applied exceeds the strength then failure will occur. An electronic component will fai ifthe applied clecrical stress (voltage, current) exceeds the ability to withstand it, and a ‘mechanical strut will bucKee ifthe compression stress applied exceeds the buck- ling strength. Overstress failures such as these do happen, bat fortunately not very often, since designers provide margins of safely. Flactrenic component specifications state the maximum rated conditions of applicstion, and circsit losignors take care that those rated valies are not excueded in servies. In most 150s they will infact do what they can to ensure thatthe in-service worst cate Seesses remain below the rated stress values this is called ‘de-ating’. Mechonical designers work in thesame way. they know the properties ofthe materais being, used (eg ultimate tensile stength) and they ensute that there Is an adequate smangin between the strength of the component and the maximum applied stress However, it might not be possible to provide protection against every possible sess application. Ifa 110V apphance is connected to a 240V supply st will obaby aan tyre wl burt if sufienty overnite lures might be cused by vurition. Inthe situations described above the values Slsarngh end hal as hesiond core lft nes eng aieege ests the known load, as shown in Fig. 12, then failure will not occur. However, in ‘most cases, there willbe some uncertainty about both. The actual strength values ‘of any poplation of components will vary: there willbe some that are relatively strong, others that aro relatively weak, but most will be of nearly average strength, Also, the loads applied willbe variable Figore 1.3 shows this type of situation. As before failure will not occur se long ab the applied load does not ‘exceed the strength. However, if there isan overlap between the distributions of lead and sirengih, and «load value it the high tail of the load distibution is applicd to an item in the weak tal of the strength distribution so that there s overlap or intererence between the distrbutions (Fig. 14), then falure will occur, We will discuss load and strength interference in more cetail in Chapter 4 4, Failures can be caused by seurout, We will use this term to include any mechan- {smor process that causes an item that is sufficiently strong atthe start of ts fe to T 3 Lowa-stengh Figure 12. Lond-stengih—dncrte values - = tes UNG Loxa-steng Figae 14 Load-trength-inering dtroione become weaker with age. Well-known examples of such processes are materi- alftigue, wear between surfeces in moving contact, corresion, irsulation deteri= ‘ration, and the wearout mechanisms of igh! bulbs snd fluorescent tubes, Figure 15 ilustrates this kind of situation. Initally the strength is adequate to withstand the applied loads, but as weakening occurs over time the strength decreases, In lerery case the average value falls and the spread of the strength distribution ‘widens, This is & major reason why it so difficult to provide accurate precic- tions ofthe lives of such items, Failures can be caused by other time-dependent mechanisms. Battery run-down, ‘creep caused by simultaneous high temperature and tensile siress, asin turbine discs and fine solder joints, and progressive drift of dctronic component param ter values are examples of such mechanisms. Failures can be caused by sneals. A sneak is 2 condition in which the system does not work properly even though every part does. For example. an elecronic system might be designed in such a way that under certain conditions incorrect ‘operation occurs The fatal iri the Apollo spacccrat crew capsule war caused in Prohabilstie Raiabilty ’ © Tmetoed es Ugicas Figure 15. Tanedepement ond and stengih vain this way: the circuit design ensured that an electrical short ciccuit would occur when a particular sequence was performed by the crew. Sneaks can also occur in software designs 7. Failures can be caused by arrrs, such 26 incorrect specifications, designs or software coding, by fauky assembly oF test, by inadequate or incorrect mainten- ance, or by incorrect use. The actual failure mechanisms that result mightinclude most ofthe list above. 8, There are many other potential causes of failure. Gears might be noisy, oi seals ‘might leak, display screens might licker, operating instructions might be wrong, for ambigueus, electronic systems might suffer from electromagnetic interfer- fence, ete Fellures have many dlifierent causes and elfets. and there are also different [perceptions of whet kinds of events might be classified as failures. The burning ‘Grring seals on the Space Shuttle booster rockets were not cassed as failures, until the dFfated launch of Challenger. We also know that all failures, in principle and almost always in practice, can be prevented. PROBABILISTIC RELIABILITY ‘The concept of reliability as probability means that any attempt to quantify it must Involve the use of statiatical methods. An understanding of statistice se applicable to reliability engineering is thorefore a necessary basis for progress, except for the special cases when reliability is perfect (we know the iter will never fail) oF itis zero (the item will never work). For practical purposes hamuner might be 109 per Cent reliable if used for driving tacks, but would havea zero reliability if used to stop ‘rain. [fused tobreak rocks, ts reliability fora given period might have some value between 0 and 100 per cent In engineering we try to ensure 10) percent redabilty, 8 Introduction t Reliability Engineering bout our expesence tells us thot we do not aways succes. Therefore relablty Satstcs ae usually concerned with probability values which are Yery high (0 ‘ery low the probability tat afalure does occu, which i I~ reliability). Quant fying such numbers brings increased uncertainty, since we need correspendingly "more information, Other sources of uncertainty ae introduced because reliability is ‘ten about people who make and people who use the prosuct, and because of the ‘widely varying environments in which typical produc might operate Further uncertainty, often ofa subjective nature, is intduced when engineers begin to discus failures. Should a faire e counted ifit was due to an eror that is hoped wil no! be repeated? I design ation is taken reduce the ska one type oF failure, how can we quantify our trust inthe designer's acca? Was the machine andor test tpieal of fhe population of machines? Reliability is quanifed in other ways, We can specify arliabity asthe mean rnamber of fire ina given Hime (Failure rate) os the mew fie bce res {GTDF for items which ave repaired. ad return! lowe, or a the me Hn fe {ature (SETTR) for ems Whi are Ret repaired. For reputed tes, 1 ofen assumed! that flues occur at 2 constant rate, in which case the failure rate AS (QITBE) | However, this Is only special case, valuable bease Iti ofen true and because tis easy to urderstar "The application and interpretation of Saistes in reliability ar less staightor- ward than in, say, public opinion pols ce measurement of human variations such as Qerheigh- in these applications, most interest i centred around the behaviour of the larger part o the population or stmple variation i not very lage ard data are Penta n relibilty we are concemed with the behaviur in the extreme tall of Aistiutions and possibly unlikely combinations of load and strength where var Shiltyisofion hard to quantily and data are expensive Farther difficulties are in application of stateBca theory t alibi: engineer ing, owing to the lat that varation is often Function of time or of tne teated fectore such as operating cydes, diurnal or seasonal cycles, maintenance periods ‘oe. Engineering, unlike moat fields of knowledge, i primarily concerned with ‘change, hopefully, but net alrays, for he better” Therefore the rliity date from any post stuation cannot be used to nuke crete foexass of the fare behaviour, without ‘aking into account ron-statstikal facts such as design changes, muirtaer Waiing, and even imponderables such eo unoresceable pro clacton or service problems. The satstcian working in relabity engineering reds to be aware of these realities. Chapter 2 provides the statistical basis of reliability engineering, but it must always be remembered that quality ard reiabilty data contain many sources of {wxcertaity and variability wich cannot be igrously quantified tis ao import- nto appreciate that fakes and thir causes are by no means alwaysclearcut and {mnambigucus They are often open to interpretation and argument. They alo difer in terms of impoctance (ont sale. ether effet) Therefore we must be carefal not te apply conventional scientific, deteminists thinkng tthe interpretation of failure, For example, a mere count of total reported failures of» producti seldom tefl or reveling It tls ue nothing about enuses or consequence, and therefore rothing about how to improve the situation. This contrasts with « satement of « ‘sical atibute such as weight or power consumpeon, whichis usually uname Repuinable an Non-Repairable Toms * biguous and complete. Nevertheless, it is necessary to derive values for decision- raking, so the mathematics are essential. The important point is that the reliability engineer or manager is no, like an insurance actuary, a powerless observer of his Satistics.Satistcal derivations of reliability arenot 2 guarantee of results, and these results can be significantly affected by actions taken by quality and reliability engineers and managers REPAIRABLE AND NON-REPAIRABLE ITEMS 1t js important to distinguish between repairable and non-repairable items when predicting or measuring reliability For a non-repairable tem sich as a light bulb, a transistor, a rocket motor or an ‘unmanned spacecraft, reliability i the survival probability over the item’s expected life or for a period during isle, when only one fire can cecur. During the iter’ ite the instantaneous probability ofthe fistard only failure is calle the ezard rat. Life values such as the mean life or mean time fo failure (MITE), or the expected life by hich a certain percentage might have failed (say 10 per ceat), are other reliability characteristics that can be used. Note that non-repairable items may be individual parts (light bulbs, transistor) or systems comprised of many parts (pacecraft microprocessors). When a part fails ina non-repairable gystom, the system fails (usually) and system reliability is, therefore, a function ofthe time tothe fist part failure. For items which are repaired when they fail, reliability is the probability that failure wil not occur in the period of interest, when more thar ene fare can occur. I can also be expressed as the fire rate or the tate of ecurrence of alures (ROCOF). However, the failure rate expresses the instantaneous prebabilty of falure per unit time, when several failures can occur ina time continuum. I is unfortunate thatthe ‘expression fatiure rate has been applied loosely in the past to non-repairabe items ‘What i really meant i that, ina repairable system which contains part type, the ‘part will contubute by that amount to the system failure rate. The pat, being non- repuirable, cannot have s failure rate. This distinction does not always have practical significance, but it can, as will be shown later. The expressicn ROCOF is sometimes used fo stress this point Repairable system reliability can also be characterized by the moar tine Ixteen Jaies (MTBF), but only under the particular condition of a constant failure rate We are also concerned with the sovilebilty of repairable items, since repai takes Hime. Availabilty & affected by the rate of occurrence of fares (ilar rat) land by maintenance time. Maintenance can be corrective (Le. repair) or preventive (to reduce the likelihood of failure, og lubrication). We therefore need to uncer: stand the raionship betwoon reliability and maintenance, and how both reliability land maintainability ean affect availabi ‘Sometimes an iter may be considered es both repairable and non-repeieable. Foe example, a missile isa rpuirable system whilst t i in store and subjected to scheduled fests, but it becomes a nen-repairable system when it is launched. Reli- ability analysis of such systems must take acccunt of these separate states. Repalt- ability might also be determined by other considerations. For example, whether w Introction to Reliability Engineering ‘TV clectronics board is treated as a repairable item or not will depend upon the cost of repair. An engine or vehicle will be treated as repairable oaly up © 3 certain age. ‘Repairable system reliability data analysis is covered in Chapter 12 and availa ity and maintainablity in Chapter 14 ‘THE PATTERN OF FAILURES WITH TIME (NON-REPAIRABLE ITEMS) “There are three basic ways in which the pattem of fares can change with time. The hazard rate may be decreasing, increasing or constant. We can tell much about the causes of failure and about the relubility of the item by appreciating the way the hazard rate behaves in time. ‘A constant hazard rate is characteristic of failures which are caused by the application of loads in excess of the design strengih, at a constant average rate. For example, overstress failures due to accidental or transient circuit overioed, or rmaintenance-induced failures of mechanical equipment, typically occur randomly and ata generally constant rate ‘Material fatigue broxght about by strength deterioration due to cyclic loading ie a failure made which doos not occur fora finite time, and then exhibits an inereasing, ‘probability of currence. Decreasing hazard rates are observed in items which become less likely to fail ‘their survival time incremes. This is oflen observed in cestronic equipment fend par. ‘Burmin’ of clectonic parts ba good exsmple of the way in which knowledge of a decreasing hazard rate is used to generate an improvement in ‘liability. The part are operated under failure-proveking stress conditions for a lume before delivery. As substandard parts fail and are rejected the hazard rate decreases and the surviving population is more reliable “The combined effect generates the so-called tathtub cure (Figure 1.6). This shows ‘an inital decreasing hazard rate or infant mortality period, an intermediate useful fe period and a final trarout period. Death is a good analogy to failure of a non- epaable sytem, andthe Eth curve model is sma fo atari statistical ‘THE PATTERN OF FAILURES WITH TIME (REPAIRABLE ITEMS) The failure rates (or ROCOF) of repairable items can also vary with time, and {important implications can be derived from these trends. ‘A constant failure rate (CFR) i indicative of externally induced failures, asin the constant hazard rate situation for non-repairable items. A CFR is also typical of complex systems subject 1 repair and overkaul, where different parts exhibit ‘ifferent patterns of falare with time and parts hive different ages since repair or rephcement. Repairable systems can show a decreasing failure rate (DFR) when ‘aliability is improved by progressive repair as defective parts which fail relatively ‘The Development of Reisblty Engiseeing 0 Flue ve tons ‘nar rary Teaure Wasa Figure 16 The ahtab’ curve carly ace replaced by good parts. “Burn in” ic applied to electronic systems, as well as to pars, for this purpose An increasing failure rate (FR) occursin repairable systems when wearout failure modes of parts begin to predominate “The pattecn of failuces with time of repairable systems can lao be illustrate by use ofthe bathtub curve (Fig. 1.6), but with the failure ate (ROCOF) ploted against age instead of the hazard rate, "The statistical treatment of failure data i covered in Chapters 2 and 3. ‘THE DEVELOPMENT OF RELIABILITY ENGINEERING Kelibilty engineering, as a separate engineering discipline, originated in the United States during the 1980s. The increasing complexity of military electronic systems was generating faluze rates which resulted in greatly reduced availabilty and increased costs. Solid state electrics technology offered long term hope, but ‘conversely miniaturization was to lead to propoctionately greater complexity, which ‘fet be relablty impeovernentsexpertel The gathering pce of earns: cave technology meant thatthe developers of new mulitary systems were making increas- ing use of large numbers of new component types. involving new manufacturing processes, with the inevitable conseqtences of low reliability. The users of such ‘equipment were also finding thatthe problems of diagnosing and repairing the new complex equipment were seriously affecting its availablity for use, and the costs of spares taining and other logistics support were becoming excessive. Against tis background the US Department of Defense and the electronics industry jointly set up the Advisory Group on Reliability of Electronic Equioment (AGREE) in 1952 The "AGREE report concluded that, to break out ofthe spiral of increasing development ‘and ownership costs cue t low rehatilityslisciplines must be laid down as integral ‘activites in the development cycle for elecronic equipment. The report laid particu lar stress on the need for new equipments to be tested for several thousand hours in n Introdetion to Reliability Ensinering high siress cyclical enviconments including high and low temperatures, vibration and switching, in order to discover the majority of weak aeas ina design at an early ‘enough stage to enable them to be corrected before production commenced. Until that time, environmental tests of tens of hours duration had been considered adequate to prove the suitability of a design. The report also recommended that {formal demonstrations of reliability, in terms of statistical confidence that specified [MTBF had been exceeded, be instituted as a condition for acceptance of equipient by the procuring agency. A large part of the report was devoted to providing detailed test plans for various levels of statistical confidence and environmental ‘conditions. ‘The AGREE report was accepted by the Department of Dofonse, and AGREE ly became a standard procedure, Companies which invested in ive environmental tet equipment necessary soon found that they could attain levels of reliability far higher than by traditonal methods It was evident that designers, particularly those working at the fringes of advanced technology, could rot be expected to produce highly reliable equipment without itbeing subjected t0 ‘a test regime Which would show up Weaknesses. Complex systems and ‘the components used in them included too many varables and interactions for the human designer to cope with infalibly, ard even the most careful design re views and disciplines could not provide sufficient protection. Consequently it was necessary to make the product speak for itself, by causing it to fail, and then to eliminate the weaknesses that caused the failures. The Department of Defense (DOD) reissued the AGREE report on testing as US Miltary Standard (MIL-STD) ‘781, Reliability Qualification and Preduction Approval Tes Meanishile the revolution in electronic device technology continued. led by integrated microcircuitry. Increased emphasis was now placed on improving the quality of devices fied to production equipments. Screening techniques, in which devieos are temperature cycled, vibrated, centrifuged, operated at electrical foverstress and othenwite abured, were introduced in place of the traditional ‘sempling techniques, With component populations on even single printed circuit boards Eecoming 30 large, sampling, no longer provided suificent protection against the production of detective equipment. These techniques were formalized, in miltary standards covering the full range of electronic cempanents, Compor tents produced to these standards were called “Hive! components, specications land test systems for electronic components, based upon the US Military ‘Standards, have been developed in the United Kingdom and in continental Europe, and internationally through the International Flectrotechnical Commission (EC). Engineering reliability effort in the United States developed quickly, and the AGREE and reliability programme concepts were adopted by NASA and many ‘other major suppliers and purchasers of high technology equipment. In 1955 the DOD issued MIL-STD-785-Refaility Pragiams jor Systems and Equipment. This docw- ‘ment made mardatory the integration of a programme of reliability engineering sctivities with the traditional enginzering activites of design, development and ‘Production, as it was by then realized that such an integrated programme was ths ‘Only way ensure that potential reliability problems would be detectod and ‘cminated at the earliest, and therefore the cheapest, stage in the development The Decwlopment of Reliability Engineering 8 cycle. Much written work appeared cn the cost-benefit of higher reliability, to sbow that effort and resources expended curing early development and during produc tion testing, plus the impostion of demonstrations of specified levels of reliability to MILSTD-761, led to reductions in in-service costs which more than repaid the reliability programme expenditure. The concept of life cycle ests (LCO), or whole Tite casts, was introduced Inthe United Kingsom, Defence Standard 00-0, The Mavagement of Reiaility and ‘Maintainebiity wasiaaucd in 1981. The British Standard Institution issued BS 576)— Guide om Reliability of Systm, Equipments and Components ‘Starting inthe early 1580s, the reliability of nev Japanese industrial and commer. far in excess of previous experience. These products were also ess expensive and ‘often boasted superior featares and performance. The Japanese quality revolution’ had been driven by the lessons taught by American teachers brought in to help Japan's post-war recovery. The two that stand out were J. 8 Juran and WV. Edwards Deming (Reference 9}, who taught the principles of ‘total quality management” (TQM) and continuous improvement. Japanese pioneers, particularly K. Ishikawa, also contributed. These ideas were all firmly rooted in the teaching of the American ‘writer on management, Peter Drucker (Reference 10), that people work most effect- ively when they are given the knowledge and authority to identify and implement improvements, rather than boing expected to work to methods dicated by ‘man agement” "eTheve ideas proved tobe mich move fective than those prescbed in documents sch asthe standards mentioned above. They led to great increaros in productivity land quality, and ths in markct penetration, as Drucker had predicted. Many Westem companies followed the new path that had been blazed and also made ‘great improvements. However, the results ofthe Japanese head start are stil appar ‘ent today in showrooms, shops, offices and facones. “Today we expect very high reiabilty in all of the products we use, Modern ‘ekcironic components and systems seldom ‘ail, and complex systems such as ‘domestic equipment, automobiles and aircraft are far more reliable than the much simpler systems of twenty oF thirty years ago. These improvements have been generated primarily by the application of the methods of TQM and continuous improvement, Many of the methods that had been standardized in the West have been shown to be less effective. ‘Another aspect of reliability thinking that has developed is the application of statistical methods Since reliability can heespressod as. probability, and is aFocted by variation, in principle these methods are applicable They forms the bass of most teaching and literature on the subject. However, variation in engineering is usually fof such an uncertain nature that refined mathematical and quantitative techn {gues can be inappropriate and miskading, This aspect will be discussed in later chapters, ‘Quality contrel (QC) of manafectaring processes obviously makes an essential contribution 1 the sellbilly of a preiuct. In the chapters on manufacturing (Chapter 12) anc reliability management (Chapter 15), the need to consider QC as an integral part of an overall reiablity programme will be emphasized 1 Introduction to Reliability Engineering ‘COURSES, CONFERENCES AND LITERATURE Koliablty engineering and management are now taught in engineering courses at 8 large number of universities, colleges and polytechnics, and on specialist short Conierences on general and specific reliability engineering and management topics have been held regularly in the United Staies since the 1950s and in Europe and elsewhere since the 1970s. The best known is the annual US Reliability and Maintainability Symposium (RAMS).sporsored by most of the important engineer- ‘ng, scociations and institutions inthe United States These conference proceedings contain much useful information and are often cited, Journals on reliability have also appeared; some are referenced at the end ofthis chapter. Several books have been published en the subjects of reliability engineering ‘and management; some ofthese are referenced at the end of other chapters. ‘Much of the relicbilty literature has tended to emphasize the mathematical and analytical aspects of the subject, withthe result that reliability engineering i often ‘considered by designers and others obea rather esoteric subject. This isunfertunate, Since if creates bamiers to communication. More recently there has been a trend t0 ‘emphasize the more practical aspects and to integrate reliability work into the overall, ‘management and engineering process. These aspects are covered in later chapters, ORGANIZATIONS INVOLVED IN RELIABILITY WORK Several oxganizations have been created to develop policies and methods in reliabil- ‘ty engineering and to undertake research and taining, A summary of these Is given in the book's homepage (see Preface tothe Fourth Eaition, page ©. RELIABILITY AS AN EFFECTIVENESS PARAMETER ‘With the increasing cost and complexity of many modem systems, he importance of reliability as an effectiveness parameter, which should be specified and pald for, has become apparent. For example, «radar station ¢ proces plant or an asliner mustbe available When required, and the cost of nortavailabiity, particularly if iti une Scheduled, can be very high. In the weapors field, if an antaircrat missile hasa less than 100 per cent probability of functioning correctly throughout its engagement sequence, operational planners must consider deplcying the appropriate extra ‘quantity to provide the required level of defence. The Apollo project second stage rocket was powered by six rocket motors ary five woald have provided sufficient impulse, but an additional motor was specified to cater for a possible failure of one. ‘As it happened there were no failures, and every launch utilized an ‘unnecessary’ ‘motor. These considerations apply equally jo less complex systems. eg. vending {and copying machines, even ifthe failure costs are less dramatic in absolute terms. ‘As an ffectivencse parameter, reliability can be ‘waded oi” against other para ‘tors, Reliability generally affcts availabilty, and in this context maintainability 6 ‘sho relevant. Reliability and maintainability are often related to availability by the formulas Reliability Programme Acticitiee 8 MIBF MIBF = MITR ‘where MTTR is the mean time to repair. Tis the simplest steady-state situation It {i clear that availability improvements can be achieved by improving either MTBF for MTTR For example, automatic builtin test equipment can greatly reduce diag rostic times for electronic equipment, at a cost of a slight reduction in overall reliability and an increase in unit costs, Many other parameters can be considemd fn trade-off, sch a8 weight, redundancy, cost of materials, parts and processes, or reduction in performance. "The greatest difficulty in estimating relationships for roliabilty trade-offs derives {from the fact that, whereas itis possible to estimate quite accurately such factors ‘as the cost and weight penalties of bulltin test equipment the cost of materials and ‘components, or the worth of a measurable performance paremete, the effect on ralabity cannot generally be forecast accurately, and reliability measurements can ‘at best Ee made anly within statistical limits imposed by the amount of data available. Even for quite simple systems, quantifying the separate contributions to ‘a measured change in reliability due o diferent causes isa very uncertain exercise Consequently, while there is considerable literature on the effectiveness of various reliability methods, few generally applicable estimating relationships are avaiable Selection of trade-oifs must therefore be very much a mater of experience o similar [Projects in the knowledge that wide margins of error can exist Availability — RELIABILITY PROGRAMME ACTIVITIES ‘Wat then, ae the scion tat managers and engineers can tks toluene 2 One obvious activity already mentioned is quality control the whole {ange of functons designed to eroure hat delivered products are compliant with the design. For many products, QC is sufficient to ensure high reliability, anc we ‘would not expect a match facory ofa company mast-producing simple diecastings toemploy reliability staf. In such cases the designs are simple and well proven, the ‘environments in which the products will operate are well understood and the very ‘occasional failure has no significant financial or operational effect. QC, together with craftsmanship, can provide adequate assurance for simple products oF when the risks are known tobe very low. Risks ae low when safety margins can be made very large, as in most structural engineering. Reliability engineering disciplines may justifiably be absent in many types of product development and manufacture. OC ‘sciplines are. however, essential elements of any integrated reliability programme. ‘A formal reliailty programme is necoscary whenever tho risks or costs of faire are not low. We have already seen how reliability engineering developed a8 a nemilt (of the high costs of unraliabilty of miltary equipment, and later in commer. ‘Gal application. Risks of filure veuslly increase in proportion to the number of components in a system, $0 relability programmes are eeguired for any prodct ‘whose complexity leads to an appreciable risk. ‘An effective reliability programme should be based on the conventional wisdom ‘of responsibility and authority being rested in ene person. Let us call im or her the w Introduction to Reliabitity Engineering rclisbility programme manager. The responsibility must relate to 4 defines abject: Ie, which may be a maximum warranty cost figure, an MTBF to be demensteated 0 requirement that failure will not occur. Having an objective and the authority, bow ‘does the reliability programme manager set about hs or her task, faced a he or she {is with a responsibility based on uncertainties? This question will be answered in detail in subsequent chapters, but a brief outline s given below. ‘The reliability programme must begin at the earliest, conceptual phase of the project tis at this stage that fundamental decisions are made, which can signifi cantly affect reliability. These are decisions related to the risks involved in the specification (design. complexity. producibilty. etc). development time-scale, re Sources applied to evaluation and test, sills available, and other factors “The chortor the dovelepment timo-scle, the more important i he need, particu larly if there will be few opportunities for an iterative approach. The activites ‘appropriate to this phase ere an involvement in the assessment of these trade-offs land the generation of seiablity cbjctves. The relability staff can perform these Tunctions effectively anly if Uhey are competent contribute to the giveand-take inherent in the trade-off negotiations, which may be conducted between designers, [production staff, marketing siaf, customer representatives and finance staf. ‘As development proceeds from initial study to detail design, the reliability risks are controlled by a formal, decumented approach tothe review af design and to the ‘imposition of design rules relating to companents, materials and process selection, derrating policy, tolerancing, etc. The objectives at this stage are to ensure that known good practices are applied. that deviations are detected and corrected, and that areas of uncertainty are highlighted for farther action. The programme con- tinues through the inital hardware manulacturing and test stages, by planning and erecting tests to generate confidence in the design and by collating. analysing and fcting upon test dats, During. production, QC activities ensure that the proven design ie repeated, and further testing may be applied to eliminate weak iteme land to maintain confidence, The dats collection, analyst and action process con. linus through the production and in-use phases, Throughout the product life cycle, therefore, the reliability is assessed, fist by initial predictions based upon past ‘experience in order to determine feasibility and to set objectives, then by refining the predictions as detail design proceeds and subsequently by recording perform lance during the test, producton and inuse phases. This performance is fed back to ‘generate corrective action, and to provide data and guidelines for future products The elements of a reliability programme are outhned in documents such as US MIL-STD-785, UK Defence Sandard 00-40 and British Standard 576) (References 1-3) The activities are described fully in subsequent chapters. RELIABILITY ECONOMICS AND MANAGEMENT Coviowsly the relisblity programme activities described can be expensive. Figure 117 ie a commonly-descrbed representation of the theorstical cost-benefit relation ship of effert expended on reliability (or production quality) activities. However, ‘despite is intuitive appeal and frequent presentation in textbooks and teaching on ‘quality and reliability this picture is misleading. Close: thought easly uncovers the Reliability Economies and Management w Tota costs uanyPotaity Progamo cons Figere17-Rotabity and fe cycle cst (tational view) error in the picture Since less than perfect reliability isthe result of failures, all of ‘which have causes, we should ask ‘what isthe cost of preventing or correcting the cause, compared with the cost of doing nothing?” When each potential or actual ‘cause is analysed inthis way, it is nearly always apparent that ftal costs continue to reduce indefinitely as reliability is improved. In other words, all effort on an ‘effective reliability programme represents an investment, usually with a large ‘payback over a short period. The only problom is that it is not easy to quantify the programme activities such as acertain aroun! of testing, ‘on achieved reliability. However, experience shows clearly that the more realistic picture isas shown in Fig. 18. twas W. F- Deming (Reference 9) who first explained this relatorship in his teaching on production qual. The truth has been clearly ‘demonstrated By the success of the companies that have wholeheartedly adopted this teaching, ‘Achieving reliable designs and products requires 2 totally integrated approach, in- cluding design, taining, fest, production, as wel as the reliability programme activ- ities In such an approach itis dificult to separately identify and cost those activities thatare specifically devoted to reliability. as apposed tosay performance or cost. The integrated engineering approach places high requirements for judgementand engin- tering knowledge on project managers and team members. Reliability specialists ‘ust play their pasta as members ofthe tea, "There are three kinds cf engineering product, from the perspective of failure prevention: 1. Intrnsially reliable components, which are those that have high margins between their strength and thestresses that could cause failure, and which do not wear out ‘within their practicable lifetimes. Such items include nearly all electronic com- Bonet (if propery aphid, neal all mechan nermoving compenents, and all correct software 1s Introduction to Reliability Engineering ost (ounces 720% Figure 18 Rally and ite cyle costs (moder view) 2 Intrinsically unreliable components, which are those with low design mar ‘which wear out suchas badly applied components ight bulbs, tarbine Blades, parts that move in contact with others, lke gears, bearings and power drive belts, ete 3. Systems which include many components and interface, like cars, dishwashers, aircraft, et, so that there are many possibilities for failures to occur, particularly ‘across interiaces (eg. inadequate electrical overstess protection. vibration nodes ‘a! weak points, electromagnetic interference, software that contains erors ee) lis the tsk of design engineers to ensure that all components are correctly applied, that margins are adequate (particularly in relation to the posible extreme valcs of strength and stress, which are often variable), that wearout failure modes are prevented during the expected life (by safe life desi system interfaces cannot lead to failure due to interacions, tlerance mismaiches, ‘e.). Because achieving allthis on any modern engineering product is 2 task that ‘challenges the capabiltics of the very best engineering teams, tis almost certain that ‘aspects of the initial design will fall short of the ‘intrinsically reabe’ criterion. ‘Therefore we must submit the design to analyses and tests inorder to show not only that it works, but ako to show up the features that might lead to faiures. When we find out what these are we must redesign and retest, until the final design is ‘considered to meet the criterion, ‘Then the product has to be manufactured. In principe, every one should be identi cal and correctly made. Of course this is not achievable, because of the inherent Variability ofall manufacturing processes, whether performed by humans or by machines. I i the task of the manufacturing people to understand and contol ‘Variation, and toimplement inspections and tests that will identify non-conforming, product Bibtigraphy ww For many engineering products the quality of operation and maintenance also influence reliability. ‘The essential points that arise from this brief and obvious discussion of fasures are that: 1. Failures are caused primarily by people (designers, suppliers, assemblers. users, ‘maintsinors). Therefore tho achievement of reliability ix sentially 4 manage. ‘ment tak, © ensure thatthe right people, skills, toams and other resources are ‘applied to prevent the creation of failures 2. Reliability (and quality) are not separate specialist fanetions that can effectively ‘ensure the prevention of falures, They are the results of effective working by all involved. ‘3. There is no fundamental Limit to the extent to which failures can be prevented. ‘Weecan design and build for ever-increasing reliability Deming explained how, in the context of manufacturing quality there is no point ‘at which further improvement leads to higher costs, Tis is, of course, even more powerfully true when considered over the Whole product life cycle, o that effets to ‘ensure that designs are intrinsically reliable, by good design and effective develop- tment testing, generate even higher pay-ofis than improvements in production ‘quality. The ‘inizer’ (continuous improvemen!) principle is even more effective ‘shen applied to up-font engineering ‘The eration of rlialeprosucs i, erfre, primarily 2 management task. Guidance on programme management and cost is covered in Chapter 18 BIBLIOGRAPHY 1. USM $10.75: aii Proms fr Sten Egle Nona Teh no Podin crvcefpraguell Wa bec oh + {iceland Manta 0. 4 Rlash Scan St Rin Sto ems nt Comps a Sad USMICSDISI: Don of Efetieaes Terns fr Rely, Manny, H Bt Sin of Evens Tres Flay, Mani man Fairs and Sey. Avaabl fron the National Techie Inermaion Service, Spring, tent TEE Sadar 578 Chery rat Clare tng by serturncpusin Brasserie forage Gece SS ee hl qual) Fhe ta id pam” Pred es ye hake pp Aner Sct for ETE pusher Wie Dolly Oxyd Clee MT Univesity Pres 1986 (orga publiedunverhe Tis curly Fotacaya Comee Pan) ve Pe Grae fc Pack Seger Neehann (1935, Introduction to Reliability Engineering QUESTIONS Define (a fate tate, and (by hazard rave. Explain ther application tothe relat of ‘components and repairable sysiems, DScuss the plausbity ofthe "athtub curve in both contexts () Explain the theory of component faiures derived from the interaction of stress (or Joad) and strengh distributions. Explain how this theory relates to the behaviour of the component hazaté function. (©) Discuss the validity of the ‘bathtub curve’ when used to describe the failure sharactetistcs of nonrepaizable components, 3. What are the main objectives ofa relablty engineering team working on an engi ‘ering development project” Describe the important sks and experience that should be avaiable within the team. Belly lst the most common basi causes of faures of engineering predcts {tis sometimes claimed that increasing quality and reliability beyond levels that have ‘been achieved in the past Lely to he uneconomic due to the cot ofthe actions that would be necessury. Present the argument against this belief, Tustrate it with an ‘©ample from your own experience 2 Reliability Mathematics INTRODUCTION ‘The methods used to quantify reliability are the mathematics of probability and statistics, In reliability work we are dealing with uncertainty. In fact this is the case in much of modern engineering. and the probabilistic as opposed to the deterministic approach Jo engineering, problems is becoming more widely applied. As an example, dats may show that s certain type of power supply fils ‘ta constant average tate of once por 107h. If we busld 1000 sich units, and ‘we operate them for 100h, we cannot say with certainty whether any will fail in that time. We can, however, make @ statement about the probly of failure. We can go further and sta that, within specified statistical confidence limits, ‘the probability of failure lies berween certain values above and below this probab- ity. If sample of such equipments s tested, we obtain data which are caled shtisis Reliability statistics can be broadly divided into the treatment of discrete functions, continuous fnctionsand point proceses, For example a switch may ether work oF not ‘work when seleced or a pressure vessel may pass or fal a test—these situations are described by discrete functions. n reliability we are often concerned with two-state discrete systems, since equipment is in either an operational or a failed state ‘Continuous functions describe thove situations which ave govemed by a continucus ‘variable, such as time or distance travelled. The electronic equipment mentioned ‘above would heve a reliability function in this dass. The distinction between discrete and continuous functions is one of how the problem is tested, and not necessarily ofthe physics oF mechanics of thesituation. Fr example, whether oF not 4 pressure vessel falls a test may be a functon of ts age, and Its relablty could therefore be teated as a continous function, The statistics of point processes are used in relation to repairable systems, when more than one failure can occur in 2 time continuum. The choice of method will depend upon the problem and en the type of data available. Concepts such as statistical confidence are often written as s-confidence to distn- ‘gaish them from the nor-statstical meanings of the terms. This convention is ‘adopted inthe remainder of this book. a 2 Relibiity Matiematies VARIATION [Reliability is influenced by variability, in parameter values such as resisance of resistors, material properties, or dimersions of parts. Variation is inherent in all manufacturing processes, and designers must understand the nature and extent of ;possible variation In parts and processes used. They must know how to measure {and control this variation, so that the effects on performance and reliability are ‘minimized. ‘Variation also exists in the environments that engineered products must with- stand. Temperature, mechanical stress, vibration spactra, and many other varying factors must be considered ‘Slatstcal methods provide tho means for analysing, undersianding and contol- ling variation. They cin help us to create designs and develop processes which are intrinsically reliable in the anticipated ervironmen’s over their expected useful lives. ‘OF course, i is not necessary to apply statistical methods to understand every engineering problem, since many are purely determinisic or easily solved using, [past experience oF information avallabe in sources such as databooks, speciicar tions, design guides, and in known physical relationships such as Ohms lew. However, thee are also many situations in which appropriate use of statistical ‘wchniques can be very eifective in optimizing designs and processes, and for solving quality and reliability problems. Accautionary note ‘Whilst statistical methods can be very powerful, economic and effective in reliability cengincering applications, they must be used in the knowledge that variation in cengincering iin important ways diferent from variation in most natural processes, for in repetitive engineering processes such as repeated, in-contrel machining or diffusion processes. Sach processes are usually Constant in time, in terms of the nature (average, spread, et.) ofthe variation. —Distributed in a particular way, descrbable by a mathematical function known as the normal distribution (which will be described later in this chapter) Infect, those conditions ofton do not apply in engineering. For example: —A.component supplier might makea small change ina process, which results in & large change (better or worse) in reliability. Therefore past data cannot be used to forecast fature reliability, using purely statistical methods. The change might be dlliberate or accidental, known or unknown. Components might be selected according to criteria such as dimension or other ‘measured parameter. This can invalidate the snormal distribution assumption on| ‘which much of the statistical method is based. This might or might not be important in assessing the results. A process or parameter might vary in time, continuously or cyclically, $0 that “statistics derived at one Hime migh not be relevant at others Probability Concrpie a Variation is often deterministic by nature, for example spring deflection as function of force, and it would not always be appropriate to apply statistical techniques to tis sort of situation, —Variation in engineering can arise from factors that defy mathematical treatment. For example. a thermostat might fail. causing a process to vary ina difierent way to that determined by earlier measurements, or an operator or test technician sight make a mistake Variation can be catastrophic, not only continuous. For example, a parameter sich ‘as a voltage level may vary over a range, but could also go & 207. ‘These points highlight the fact that variation in engincering is caused to a large ‘extent by people, as designers, makers, operators and maintainers. The behaviout and performance of people is not as amenable to mathematical analysis and fore- casting as i, say, the response of a plant to fertilizer or even weather pattems to ‘ocean temperatures, Therefore the human element must always be considered, and statistical analysis must not be relied on without appropriate allowance being made for the effects of factors such as motivation, training, management, and the many other factors that can influence reliability. Finally, it is most important to bear in mind, in any application of siatistical ‘methods to problems in science and engineering, that ultimately all cause and effect relationships have explanations. in scientific theory. engineering design, process or hhuman behaviour etc. Statistical techniques can be very useful in helping us to understand and control engineering situations. However, they donet by themselves provide explanations. We must alteays seak to understand causes of variation, since fly then can we really be in control. Ser the quotations on the flyoaf. and think bout tem, PROBABILITY CONCEPTS ‘Any event has a probably of occurrence, which can be in the range 0-1, A zero probability means thatthe event will net cccur; probability cf 1 means tht wil Secu. A cain has a 05 (even) probsbiity of landing heads, and a cie hes 2 1/6 probably of giving any ane of the six numbers. Such events aresndpenier, ve the con nd the die logically hae no memory, 50 whatever has been thrown in the past cannot affect the probability of the next throw. No ‘system’ can beat the ‘Statistics ofthese situations waiting, fora run of blacks at roulette and then betting ‘on reds only appears to work becase the gambles who won tis way tak about ‘whist thase whe lst do not With cons, dice and roulette wheels we an predict the probability ofthe outcome from the nominal nakire of the system. A coin has two sides, a die sx face, 2 roulette wheel gual numbers of reds and blacks. Assuming thatthe coi, die and wheel are fir hove outcomes are ls alized Le. they are alleqully probable. In ther words, they cea rly ‘With many systems such a the sampling of items (fom a prestuction butch, the probatliies can only be determined from the stalsics of previous experience ‘We can define probability in wo ass: 4 Reliability Mathematics 1. Han event can cccar in N equally likely ways, and if the event with atibute A can happen in 1 ofthese ways, then the probability of A occurring is Pray= 2 N 2. I in an experiment, an event with attibute A occurs n times out of N experi- ‘ments, then as N becemes large, the probability of event A approaches 2/N, ic. Pra ‘The firt definition covers the cases described carier, Le equally likely sinde- pendent events euch a8 rolling dice. The second definition covers typical cases in ‘quality control and reliability. Ifwe test 10 items and find that 30 re detective, we ‘ray feel justified in saying that the probability of finding, « defective item in our rex test 5 0.30, oF 30 per cent However, we must be careful in making this type of asertion, The probability of, (030 of Finding a detective item in our next test may be considered as our degree of Die, limited by the size of the sample, in this outcome. This leads to a third, subjective, definition of probability. Uf, in our tests of 100 items, seven of the defectives had occurred in a particular batch of ten and we had taken corrective ‘action to improve the process so that such a problem batch was less likely to occur in future we might assign some lower probability to the nextitem being defective. This ‘subjective approach i quite valid, and is very often necessary in quality contrl and Feliabilty work. Whilst itis important to Rave an understanding of the ruies of ‘probability. there are usually so many variables which can affect the properties of Iranufactured ileme that we must always keep an open mind about statistically derived values. We must ensure that the sample from which statistics have been derived represents the new sample er the overall peptlation, about which we plan to make an assertion based upon our sample statistics. ‘A sample represents & population if ll the members of the popalation have an ‘equal chance of being sampled. This can be achieved Ifthe sample fs selected 50 that, this condition is flied. Of course in engineering this s not always practicable; e3 in reliability engineering we often need to make an assertion about items that have not vet been produced, based upon statistics from prototypes. "To the extent tha the sample s rot representative, we will alter our assertions. OF ‘course, subjective assertions can lead to argument, and it might be necessary to perform additional tests to obtain more data to use in support of cur assertions. Ii ‘we do perform more tests, we need to have a method of intempreting the rev data in relation tothe previous data: we sell cover this aspect later ‘The astertions we can make based on sample statistics can be made with a degree ‘of econfidence which depends upon the size of the sample. If we had decided 1 test tn items alter introducing a change to the process, and found one defective, we ‘might be tempted to assert that we have improved the process, from 30 per cont defectives being produced to only 10 per cent. However, since the sample is now ‘uch smaller, we cannot make this assertion with as high confidence as when we aes of Probabiity 2s eat toeocosooco8o z2o00#8c #00000 soeoco#8ooe8G soeococfcagae seococeeeoeca eeocccoeeooo Figers 21. Sample with defies (ack ques) ‘used a Sample of 100 In fact, the true probatility of any item being defective might Still be 30 per cent, i.e the population might stil contain 30 percent defectives Figure 2.1 shows the situation as it migat have occurred, over the fist 100 tests. ‘The black squares indicate defectives, of which there are 30 in cur batch of 100. If these are randomly distributed, its possible to pick « sample batch of ten which contains fewer (or more) than three defectives. In fact, the smaller the sample, the ‘greater will be the sample-to-sample variation about the population average, and the s-confidence associated with any estimate of the population average will be accordingly lower. The derivation of confidence limits is covered later in this chapter RULES OF PROBABILITY In order 10 utilize the statistical methods used in reliability engineering, itis neces: sary to understand the Easic notaton and rules of probability. These are: 1. The probability of obtaining an outcome A is dencted by P(A), and so on for other outcomes. 2. The joint probability that A aud B occur is cenoted by P(AB). 3, The probability that A or B occurs is dented by P(A + B) 4. The canditinal probability of obtaining outcome A, given that B jus occured, is denoted by P(A'B) 5. The protability of the complement ie. of A not cccurting, is PCA) 6. W (and only if events A and Bare sindependen! then, ~ PA. 2% Reliability Mathematics and PIBIA) = PBIA) = PB) ae) ie. PIA) is unrelated to whether or not B occurs, and vice versa 7. The joint probability of the occurrence of two s-independent events A and Bis the product of the individual probabilities: PAB) = PCA)PCB) ey ‘This i aloo called the product rule or serie rule. It can be extended to cover any number of: independent events. For example, in rolling adie, the probability of obtaining any given sequence of numbers in three throws is iaied ae 6 1 Wevents A and Bare sdependent, then PLAB) = PLAYPIBIA) = PCB)PCA\B) @) ce. the probability of A occurring times the probability of Boccurring given that ‘Avhas already cocurred ce vice Vers. IEP(A) # 0, Eon (2.3) can be rearranged to PAB) rea =a co) ‘9, The probability of any one of two events A or B occurring is PLA +B) = P(A) + P¢B) — PAB) es) 10, The probabilty of A or Boccurring if A and B are sindepenstent is PLA +B) = P(A) +PCB) ~ A)P(B) 26) “The derivation of this equation can be shown by considecing the system shown in Fig. 22, in which either A oe B, or A and B, mast work for tho eystom to work. If we denote the system success prebabilty as P, thon the failure prebability, Pj ~ 1 Py ‘The system failure probability isthe joint probability of A and B falling, Le a 1 ~ PAI] HL PB = 1-P(A)— PB) + KAYE) 1—Py=P(A +B) = (A) + FB) ~ PLAPCB) ales of Probability 2 Figere22 Du redundan stem 11, events A and B are mutuily exchasioe i, A and 8 cannot accu simultan- ‘oval, then PAB) =o and PIA +B) = P(A) + 4B) en 12, If multiple, mutually exclusive protabilies of outcomes B, jointly glvea prob- ability of outcome A, then PIA) = PAB) = SP(AIB PB) es) 13, Rearranging Eqn (23) PAB) = PLA)P{BIA) = PCB)PIA(B) wwe obtain PAPA) rain) = PAT) a9) This isa simple frm of Bays theorem. A more general expression is PAYPOBIA)_ PAB) = sy BPE 210) where Fs the ith event. Example 21 ‘The reliability of a missile is O85. Ifa salvo of two missiles is fred, what isthe probability of at least one hit? (Assume sincependence of missile hits) Let A be the event first misile hits’ and B the event ‘second missile hits. Then PXA) = PB) = 085 eli ‘There are four possible, mutually exclusive outcomes, AB,A\ probability of both missing, from Eqn (22), 18 PAB) 15° 00035 Paes) “Therefore the probebilty of a last ne itis y= — 0.0225 = 0975 ‘We can derve the same result by using Eqn (2.9): PA +B) = P(A) + PCB) ~ PLAYI(B) 85 +085 - 089° = 0975 Another way of deriving this result is by using the sequence tee diagram: FD) ~ AAIRO) ~ 086086 - 07885 UB) = RAP) =085%0.15= 01275 tip) = ABAD) <0.15-088 0.1275 FB) = RAR "The probability of ahi is then derived by summing the products ofeach path which leads to at last one hit. We can do this since the events defined by each path are ‘mutually exclusive, PAB) + P(AB) + PIAB) 9778 (Note that the sum ofall the probabilities is unity) Example22 In Example 2.1 the missile hits are no! independent but are s-dependent, so that if the first missle fails the probability that the second will also fal is 012. However, ifthe first missile hits, the hit probability ofthe secand missle is unchanged at O85, ‘What is the probability of at last one hit? Rates of Probability » “The probability of atleast one hit is PIAB) + P(AB) + PIBA) Since A, B anc AB are sindependent, LAB) = PLAC) = 085 «0.85 ~ 0.7225 01275 and the sum of these probabilities is 0.97. "This result can also be derived by using a sequence tre diagram: Twa hts As in Example 2.1, the probbility of at leest one hit is calculated by adding the products of esch path leading to af least ene hit, PLAYA) + PAPO) + PCAYPCD) (085 5 085) + (O85 % 0.15) + (015 x 080) 0 Reliability Mathematics Example23 In the circuit shown, the probability of any switch being closed is 08 and all events are «independent. (a) What is the probability that a cincuit will exit? () Given that a circuit exists, what isthe probability that switches a and b are closed? Let the events that a,b, cand d are closed be A, B, Cand D. Let X denote the event that the circuit exists @X=ABHIC+D) P(X) ~ (AB) + {C+D} —PIAB)NC + D) (AB) = P(AIPG) 18 x 08= 064 PIC+D) = PC)+ POD) — PIPED 5408 - 064 =0.96 Therefore PX) (4 +096 ~ (0.95 x 068) = () From Eqn (24), PIABIN) = ae A and B jointly give X Therefor, from Fan 2.8), PIABX) = PLAB) so Fyn) ABIX) a Example 24 A test set has a 98 per cent probability of correctly classifying a faulty item as fective and a 4 per cent probability of cassiiying a good item as defective I in ‘batch of items tested 3 per cent are actually defective, what i the probability that ‘when an item is classified as defective, ti truly detective? Let D represent the event that an item is defective and C represent the event that an item is classified defective. Ten 4) ~ 003, ACID) = 098 ACID) = 008 We need to determine P{D|C). Using Eqn (2.10), PDIPCID) READ +PCPD (oo3y(098) TaMH.B)+ ODO] AIC) 143 ‘This indicates the importance of a test equipment having a high probability of correctly classifying good item as well as Bad items. CONTINUOUS VARIATION “The differences, or curiations of parameters in engineering. applications (machined dimensions, material strengths, transistor gains, resistor values, temperatures, etc) are conventionally described in two ways The first, and the simplest, is to state ‘maximum and minimum values, o tolerances, This provides no information en the nature, or shape, of the actual distribution of values, However, in many peoctical cases, this is sufficient information for the crestion of manufacturabie, reliable designs. “The second approach isto describe the natare of the variation, using data derived from measurements. In this section we will describe the methods Of statistics in relation to deserbing and controling variation in engineering. Ifwe plot measured values which can vary about an average (eg, the diameter of ‘machined parts or the gain of a transistor) 2s histogram, for a given sample we ‘may obtain a representation such as Fig. 23a). Tin this case 30 items have been measured and the frequencies of occurrence ofthe ‘measured valuesare as shown, The values range from 2t09, with mostitems having ‘aloes between § and 7. Another rancom sample of 8) from the same popwation ‘will usually generate a diferent histogram, but the general shape is likely to be similar, eg Fig. 23) I we plot a singe histogram showing the combines dats of 2 Reliability Mathematics ol* ol ssf Pa pe eo de. be | Ba fa iE | teraseves = 7a Tee Taaeseres ‘ake ‘vane Figure 23_ (2) Frequency histogram of a rardam sample, (b) frequen histogram of another ‘tds snp on ane ppl) dita may sapien tr wih ene ‘many such samples, but tis timeshow the values in measurementinterval of 05,We get Fig. 23(9). Note that now we have used a percentage frequency scale. WVe Now havea rather beter picture of the distribution of values, as Wehave more informaton, from the larger sample. If we proceed to measure a large number and we further reduce the messurement interval, the histogram tends to acurve which describes the population probability density function (pdf) or simply the distribution of valves. Figure24shows a general unimodal probability distribution, fx)being the probability density of occurrence, relate tothe variable x. The value of t which he distribution ‘peaks is called the mode. Multimodal distributions areercountered in reliability work wel as unimodal distributions. However, we will deal only withthe statistics of ‘unimodal distributions in this book. since multimodal distributions are usually {Generated by the combined effects of eparate unimodal distributions, "The area under the curvels equal tony, since t describes the total probability of all possible values of s,s we have defined a probability which is a ertsinty as beng, 8 probability of ene. Therefore [ters eu ‘The probability of a value falling between any two values xy and ay is the area bounded by this imerval, i 4) dx (212) Figee24- Continuo probably dibuton Continnons Veriation 2 To describe a pdf, we normally consider four aspects: 1. The ceatriliendency, about which the distribution is grouped. 2 The spreaf, indicating the extent of variation about the central tercency. 3. The skewness, indicating the lack of symmetry about the central tendency, 44. The kurtesis. indicating the ‘peakedness’ of the pl ‘Measures of central tendency For sample containing m items the sample mn is denoted by oe (213) ‘The sample mean can be used to estnale the population mean, whieh isthe average ‘of all possible outcomes. For a continuous distribution, the mean is derived by ‘extending this idea to cover the range ~ to +30 ‘The mean ofa distrioution is usually denoted by jThe mean Isalso referred toas the location parameter, average value or Sexpeted Tal, E(). In reliability Work #15, used to denote the mean in certain special cases (see page 3) o-[ soe any ‘This is analogous to the centre of gravity of the pul. The estate of a population ‘mean irom sample datais denoted by ji. Other measures of central tendency are the median, which isthe mid-point of the distribution ie. the point at which half the measured values fall toeithe: sce, and the mode. which is the value (or values) at which the distribution peaks. The relationship between the mean, median and mode for a rightskewed distribution is shown in Fig. 25. Fora symmetrical distribution, the three values are the same, ‘whit for a leltakowed ditrbntion the onder of values i reversed. Spread of a distribution ‘The spread, oF dispersion, ie, the extent to which the values which make up the slstibution vary Ismeasured by Its variance, For asamplesize nthe valance, Var (2) for E(x~ 7, Isgiven by o Reliability Mathenaties Saw Vari) = Ete 3 = a5) ‘Where sample variance used to estimate the popslaion variance, we ase (#—1) jn the deneminator of Ean (215) instead of nat can be show to provide a tattor ostmate Tho rtnte of pepuaton taro from a sample + devoted o FS a6 ‘The population variance o, for a finite population N, is given by _ Sei oan ain fer ordre ibe of a-wr aos 2) rin called the sanderd deviation (SD) and ie frequently wed in practice instead of the ‘ariance. Its also referred to as the sale paremster. 2 i Uhe second moment sbout the mean and is analogous toa mdius of gyration. Skewness and kurtosis ‘The third ard fourth moments about the mean give the skewness and kurtosis of the lstribution. Since we will not make use ofthese parameters in this book, the reader {sreferred to more advanced statistical texts for their derivation (es. Reference). A positively stewed distribution is skewed to the right, and vice versa. Asymmetrical ddstribution has zero skew. ‘The cumulative distribution function ‘The cumulative distribution function (cat), F2), gives the probability that a meas- ‘ured value will fall between oe and 3: noe [ see a) Figure 26 shows the typical ogive form of the edlf. with F(a) + Las x Contino Distribution Pnctions s car Igere 29 Typical cimtlatve dntbuton tinction eat) Reliability and hazard functions In reliability engineering we are concerned with the probability that an item will survive fora stated interval (eg. time, cycles, distance, etc), ie. that there i n0 failure in the interval (49 3). Ths is the , The lognormal Pal es Reliability Matheraties 1 1 ins» =ree|-} for s20) 4) =) man | 5( | aaa em) ° (for <0) Inother words it is the snormal distribution with Inx as the variate. Te mean and SD of the lognormal distribution ae given by senean() SD = [exp Qe +20") —expQuet ol? ‘where and o are the mean and SD of the In data. When j: > ¢, the lognormal distribution approximates to the s-normal distibu- tion. The ¢normal and lognormal distributions describe reliability situations in ‘which the hazard rate increases from x = 0 to a maximum and then decreases. “The exponential ditibation ‘The exponential distribution describes the situation wherein the hazard rate is constant. Poisson process (page 71) generates a constant hazard rate The pal f i eae This is an important distribution in reliability work, as it has the same central limiting relationship to life statistics 26 the rormal distribution has to non-ife statistics It describes the constant hazard rate situation. As the hazard rate is often function of time, we will denote the independent variable by {instead of x. The constant hazand rate is denoted by A. The mean life, or mean time to failure (MITE), is 1/A. The pa. is then written as “ dep A) (2.26) “The probability of no failures occurring before time tis cbtained by integrating, Eqn (226) between 0 and # and subtracting from 1 RD) [eae e0t—10 an Rit) isthe riabilty function (or survival probability) For example, the reliability of an item with an MITF of S00h over 2 2th period is G4) Continnonn Disivintion Factions » Values of exp( x) are tabulated in Appendix 2. Note that for sonal values of MC < 0.1), the first two terms of the negative exponential series (1 At =...) can ‘be used to give a good approximation for RC). Rt) can alsobe derived a follows. If there are Nitems surviving a time t then in a small period of time 6 the number failing is ON. Then a = -AN {i.e constant proportion to number surviving aN Wow -N=Inc Copia TFN i the numberof items at O then ¢ For tems which arerepaired, \is called the ature rate and 1/Aiscalled the mean tine ‘beter furs (TBF) (6). Note that €3.2 percent of items will have fled byt = 9 If times to failure are expenentally distributed. the probability of x falures is Poisson-distributed (Page &5) Far example, ifthe MTBF is 100h, the probability of Ihaving more than 15 failures in 1000 is derived as: [Expected numberof fatlures =: 700 Referring to Fig. 2.17. for Nui) = 10. x P0965 ‘This the probability of having up to 15 failures. Therefore the probability of more than 15 failures is 10.965 = 0.035 ‘The gamma distribution ‘The gamma distibution describes i roliabilty terms, the situation when parti failures can exist, Le. when 2 given number of partial flare events muct occar before an item fils, or the time to the ath failure when time to failure is expenen- tilly dsteibuted. The pa. is 0 Relibiity Mathematics ww-{ ¥ ae AyQateapt =A) or x20) @ o Gor x <0) (228) ; he tala ope ie and te unter asp SAR Oe ra)=[eMeptnas ax ‘When (a ~ 1) is positive integer F(a) = (a~ 1)! Thisis the casein the partial failure situation. The exponential distribution isa special case of the gamma distribution, when «= Tie. fe) = Soap 0) ‘The gamma distribution can also be used 19 describe a decreasing or increasing hazard rate, When 2 < 1, hix) will decrease whilst fer a > 1, h(x) increases. ‘The x’ distribution ‘The x (chi-square) distribution is a special case ofthe gamma distribution, where A= [and v= 2/2, where» is called the number of degrees offredom and must be a poslive integer. This permits the use ofthe x? distribution for evaluating reliability Sisetions, sine the numberof flares, of events to file, will sways be positive Integers. The 1° distribution is realy a family of distnbutions, which range in shape {rom that ofthe exponential wo thatof the social dstibution, Bach distabution is ‘dentified by the degrees of freedom. Instatistical theory, the x distribution is very important, as itisthe distribution of the sums of squares of », independent, normal Variates. This allows it tobe used for statistical testing, goodness-of-fit tests and evaluating s w-fererb(] se ne ° (ort <0) ‘The corresponding reliability inction is no =e[-(9)| em Bis the shape parameter and is the sole parameter, ox characteristic Hfe—it is the hfeat ‘which 632 per cent of the population will have fale. ‘When f= 1, the exponential reliability fanetion (constant hazard rat) revue, vith y= mean life (1/9). ‘When p< 1, we get a decresing hazard cate eiablty function ‘When f > 1, we get an incrasng hazard rate reliability function. ‘When f = 35, for example, the distritution approximates to the normal distibu- tion. Thus the Webull distribution can be used to model a wide tange of lfe istibutions characteristic of engineered products So far we have dealt with the two-parameter Weibull distribution. If, however, failures do notstartat = 0, but only after 2 finite time», then the Weibull reliability function takes the form wo meol-(52)] om i.e a three-parameter distribution.» is called the failure fre tie, locaton parameter or ‘minima Tif Tis xomsoties denoted as te "Another parameter that is used is the Buife. which isthe time by which 3% ofthe population can be expected to fail. tis expressed ae B,, where ns the proportion failing, ‘The extreme value distributions In reliability work we are ofien concerned not with the distribution of variables Which describe the bulk of the population but only withthe extreme values which 2 Reliability Mathematics ‘Table 24. Sample data taken randomly from 4 commin popltion| Sumple Data 1 2 a Bw 0 2 a 5 a a Ss 2 ie oe u 4 BS at 0 5 i 2 2 2 & a ‘ 2 %» @€ F 8 m 7 > s 2B » B® 2 5 SN a ee . x = * S&S S 8 S&S BF aw a ea or ee u > & 2 b Ss e & A 2 Se ae ae a oF can lead to failure For example, the mechanical properties of a semiconductor wire bond are such that under normal operating conditions good wire bends will not fracture or overheat. However, extreme high values of electrical land or extreme low ‘values of bond strength can result in faire. In other words, we ae concerned with the implications ofthe tale of the distributions in load-strength interactions. How ‘ever, we offen cannot assume that, case 3 measured value appears to be, sy. #- normally distributed, that this distribution necessarily is a good model for the extremes. Also, few measurements are likely to heve been made at these extremes, [Extreme value statistes are capable of describing these situations asymproticlly Entreme value statistics are derived by considering the lowest or highest values in each ofa series of equal samples. For example, consider the sample data in Table 2.1, taken randomly from acomman population. The overall data can be plotted as shown in Fig. 29 as #12). However, if we plot separately the lowest values and the highest vvaluesin each sample, they will appear as g(x) and gy(s)g1(2)isthe extreme value distribution of the lowest extreme whist g(x) isthe extreme value distribution ofthe highest extreme in each sample. For many distributions the distribution of the extremes will be one of three types: Type I —also known as the extreme value or Gumb! distribution ‘Type Hl —also known as the log extreme value distribution. ‘Type Ill —for the lowest extreme values. Ths is the Weibull distribution Extreme ealue type I ‘The type I extreme value distributions for maximum and minimum values are the limiting models forthe right and left tails ofthe exponential types of distribution, where ths is defined as any distribution whose cumulative probability approaches ‘unity ata rate which is equal to or greater than that forthe exponential distribution, ‘This includes most reliability distributions. such 26 the normal, lognormal and ‘exponential distrbutions ‘The probability density functions for maximum and minimum values, eespect- ively, re Continsone Distribution Functions 2 tox. Figare29. Exveme value distributions fo tea{ ten oof tee-a]} on) £09) =See{ Fer e[26-n]} es) ‘The reducs onriate i given by ‘Substitating in Eqn (239) and (234), wean derive the efi terms ofthe reduced variate y. For maximum values Fy [p(t +e(—mide=ept—exp(-y) 235) “ Reliability Mathematics For minimum values Fu)=1-expl expel (230) The distribution of maximum values is right-skewed and the distribution of min- imum values i left-skewed. The hazard function of maximum values appreaches unity with increasing x, whilst that for minimum values increases exporentaly. wand 9 are scale and location parameters.» isthe mode of the distribution. «and ‘pare related to the mean and standard deviation ofthe distributions of the extrome valuonby. ‘Malm values Hane = 2+ 057TH Minieisn vateos: evag = P0577 ‘The standard deviaticn soy is 1.2831 in both cases. Extreme value type I ‘The extreme type Il distribution does not play an important role in reliability work. Ifthe logarithms ofthe variables are extreme valve distributed, ten de variable is slescriec by the extreme value type Il distribution. Thus its relationship tothe type feteme value distribution 4s analogous to that of the lognormal to the normal distribution. Extreme zlue type Ut ‘The type Il extreme vale distribution for minimum values ie the limiting ‘model for the leftard tail for distributions which are bounded atthe left. Ia fact the Weibull distribution is the type Il extreme value distribution for minimam values, and although it was initally derived empirically, its use for describing the strength distribution of materials has been justified using extreme value theory. ‘The extreme value distributions related to load and strength ‘The type | extreme value distribution for maximum values is often an appropriate ‘model for the occurrence of load events, when these arenot Founded to the right, Le ‘when there is no limiting value, tis well known that enginevring: materials possess strengths well below thei theoretical capacity, mainly die tothe existence of imperfections which give rise 0 nontiniform streseos under load. In fat, dhe strength will ho rlated to tho effect of the imperfection which creatos the greatest reduction in strength, and honco the Verition in Enginesing 6 extreme value distribution for minimum values suggests itself as an appropriate ‘model for strength, ‘The strength, and hence the time to failure, of many types of product can be ‘considered to be dependent upon imperfections whose extent is bounded, since only very small imperfections will escape detection by inspection oF process canta justifying uce of a type III (Weibull) model. On the other hand. a type | model ‘might be more representative of the strength of an item which is mase-preciced an ‘not 100 per cent inspected, of in which defect can exist whose extent is not bounded, but which are not detected, eg. a long wire, whowe strength wil then be a function of length. For a system consisting of many components in series, where the system hazard rateis decreasing from £= 0 (Le. bounded) a type Ill (Webull) distribution willbe 2 {good mode! for the system tie to failure, SUMMARY OF CONTINUOUS STATISTICAL DISTRIBUTIONS. Figure 2.10 is 4 summary of the continuous distributions described above. VARIATION IN ENGINEERING Every practical engineering design must take account of the effects of the variation inherent in parameters, environments, and processes. Variation and its efiects can be ‘considered in three categories: 1. Deterministic, or canal, which isthe case when the relationship between a param- ter and its effect Isknown, and We can use thecretscal or empirical formulae, 8 ‘we can use Ohm's law to calculate the effect of resistance change on the perform ance of a voltage divider. No statistical methods are required. The effects of variation are cau by inserting he expected range of vals ino the formu 2. Functional, which includes relationships such as the effect of a change of operat- ing procedure, human mistakes, calibration errors etc. There are no theoretical formulae. In principle these car be allewed for, but often are not, and the cause and effect vationchips are not always easy to identify or quantify. 3. Renlom. Those are the effects ofthe inherent variability of processes and operat- ing conditions. They can be considered to bo the variations that are left unex plained when all deterministic and functional causes have been liminated. For example, a machining proces that is in control will nevertheless prod luce purts with some vasiaticn in dimensions, ard random voltage fluctuations ‘can occur on power supplies due to interference. Note that the random varat= fons have causes. However, i is not always possible or practicable to predict ‘now and when the cause will arse, The statistical models described above can ‘be used to describe random Variations, subject © the limitations discussed late, rs Reliability Matheratcs P| ou 8 (oo unger ut unas suonsuny aes pazey pur SagernH SusgNgUSP une wouD Jo sdeys OL AMSA [leew Weniaion te Engincring eats acts es (i Sat} 8 c- ont ol Yo Sem ey wonou ie _ an a Reliability Mathematics Variation can also be progressive, for example due to wear, material ftigue, change of lubricating properties oF electrical parameter drt {the variation -normal? ‘The central limit theorem, and the convenient properties of the s-normal distribu- tion, are the reasons why this particular function is taught as the basis of nearly all statistics of continuous variation. It is a common practice, in most applications, to assume that the variation being analysed is s-normal. then to determine the ‘mean and SD of the snormal distribution that bes its the data However, at this point we must stress an important limiation of assuming that ‘the «normal distribution descrives the real variation of any process. The enormal pdt has values between 4 and ~2o. Ofcourse e machined componentdimension Sennot vary lke this. The machine cannet add material tothe component, 30 the dimension of the stock (which of course will vary, but not by much) will set an ‘upper limit. The nature of the machining process, using gauges ox ater practical limiting fearures, wil set a lower limit. Therefore the vanation of the machined dimension would more realistically look something like Fig. 21. Only the central [part might he approximately snormal, andthe distribution will have been curated ‘In fact al variables, whether naturally occurring or resulting from engineering or ‘other processes, are curtailed in some way, so the -rormal distribution, while being ‘mathematically convenient, is actually misleading when used to make inferences ‘well beyond the range of actual measuremen's, suchas the probability of meeting an adult who is 1m tall Prbbily : x sind devatons Figure 21 Cred normal dtuton Varieties a Eninaving ° Tam Non oS sIOm Papel igure 212, ctf wc There are other ways in which variation in engineering might not be s-normal. These Thor might be othor kinds of selection process. For example, when electroaic ‘components sich as resistors, microprocostors, ete ate manusactied, they areal tested at the end of the production process. They are then ‘binned’ according t0 the measured values. Typically, rotors that fall within 12 per cen’ of the nominal resistance value are classified as precision resistors, and are Inbelled, ‘binned and sold as such. Those that fall outsice these lita, but within £10 per ‘cent become non-preision resistors, and are soll ata lower price Those that fall ‘outside +10 per cent are scrapped. Because those sold as +10 per cent will not include any that are-+2 per cent, the distribution of values s&s shown in Fig. 212. ‘Similarly, microprocessors are sold as, say, 200 MHz, 400 MHz, S00MHz, etc ‘operating speeds depending on the maximum speed at which they furction| cecrectiy on test, having all been produced on the same process. The different ‘maximum operating speeds are the result ofthe variations inherent in the process of manufacturing millions of transistors and capacitors and their interconnections, ‘on each chip on each wafer. The technology sets the upper limit forthe design and the process, and the selection criteria the lower limits. Of course, the process will alco pmodice a proportion that wil not mest other aspects ofthe specification, oF | that will not work stall Tho variation mightho unsymmetrical or sense, aschown in Fig 213, Distribution functions such asthe lognormal and the Weibull ean be seed to model uneymmet real varistion. However i sil important to remember that these mathematical ‘models will still represent only approximations to the true variations, and the farther into the tails that we apply them the greater wil be the scope for uncertainty and error The Variation might be multimodal (Fig. 214), rather than unimodal as repre- sented by distibution functions like the “normal, lognormal and Web func- tions, For example, a process might be centred on one value, then an adjustment 50 Reliability Mathematics Prebabity Figere 212 Shawed datrution Sonne! mena 10 MW 6 RIOT Figere 24 Bi modal datibution moves this nominal value, In a population of manulactured components this ‘might result in a total variation that has two peaks, or « bemodal distribution, ‘Acomponent might be subjected to a pattern of strew cycles that vary over a range in typical applications, and a further stess under particular conditions, for ‘enample resonance, lighting strike, etc Variation of engineering parameters is. to a lange extent, the result of human performance. Factors such 3s. measurements, calibrations, accept/rejoct criteria, Control of process, et. are subject to human capabilities, jadgements, and errors People do not behave s-narmally ‘Walter Shewhart, in 1931 (Raference 3) was the frst to explain the nature of variation in manuacuting processes. Figure 215 illustrates four very different kinds of variation, which, however all have the same means and SDs. These show ” Figure 225. Four datiutene with he tame manne ar SD (aher WA. Showhact) clearly how misleading itcan be i assure that any variation is “normal and then to ‘make assertions about the population based upon the assumption. fects and causes Inengineering (and in many other applicators) we are zeally much more concerned ‘withthe effects of variation than with the properties and perametess. If forexarple, the output of a process varied as in Fig. 213(0), and the “ine” lines denoted the allowable tolerance, 100 per cent would be in tolerance, If, however, the process behaved as in (a) oF (4), proportion would be outside tolerance (ony atthe high lend in the case of (4). Variation can have other effects. A smaller hameter ona shaft ‘might lead to higher ol loss oF reduced fatigue life. Higher temperature operation ‘might make an electronic circuit shut down. A higher proportion of fast processors on. ‘production batch would result in higher profit. We must therefore fist identify the effects of variation (they are often starkly apparent) and determine whether and to ‘what extent the effects can he reduced. This isnt simply a matter of ‘reducing SD’. “The eficis of variation can be reduced in two ways: 1. We can compensate for the variation, by methods such as gauging of ‘select on test (Chi curtail the original variation), by providing temperature compensating vices, ete 2. We can reduce the variation. Inboth cases we must understand the cause, or causes ofthe Variation. Shewhart categorized manufacturing variation into ‘asigneblé and ‘now-essignabe™ causes. 2 Reliebility Matienaties (These are also referred to as ‘special cuss” and ‘orion causes) Assignable varie ation is any whose cause can be practically and economically identified and re- ‘ced: deterministic and functional variation fal into this category. Non-assignable vvanation ig that which remains when all of the assignable variation has been removed. A process in this state is ‘in control, and will have minimal, ranciom, variation. Note that these are practical criteria, with no strict mathematical ba Shewhart developed. the methods of statistic! proves control (SPC) around this thinking, with the emphasis on using the data and charting, methods to identify and reduce assignable variation, and to keep processes in contol. SPC methods are described in detail in Chapter 13, Trails People like life insurance actuaries, clothes manufacturers, and pure scientist ere interested in averages ard SDS: the behaviour ofthe bulk ofthe dat, Since most of the sample data, in any situation, will represent this behaviour, they can. make credible assertions about these population parameters. However, the farther we try to extend the assertions into the tails, the less credle they become, particularly ‘when the assertions are taken beyond ary of the data. In engineering we are usually more concerned about the behaviour of variation atthe extremes, than that near the avenge. We are cancemed by high stresses, high and low temperatures. slow ‘processors, weak components. ec. In other words, itis the tals of the distributions that concern us. We often have only small samples to measure or test. Therefore, "using conventional mathematical statictics to attompt to understand the nature, uses and effects of variation at the extreme can be misleading. However, thene Situations can be analysed using extreme value methods (page 41) Conclusions ‘These are the aspects that matter in engineering, and they transcend! the kind of basic statistical theory that is generally taught and applied. Laier teachers, particu~ larly W. B. Deming (Chapter 1, Reference 9) and Genichi Taguchi (Chapter 5) cestended the ideas by demonstrating how reducing variation reduoss costs and ‘increases productivity, and by eonphasizing the management implications. Despite al of these reasons why conventional statistical methoxs can be mislead: ing if tse to describe and deal with variation inengineering, they are widely taught and used, and ther limitations are hardly considered. Examples are: Most textbooks and teaching on SPC emphasize the use of the normal distribu tion as the basis fr charting and decision-making, They emphasize the mathemat- ical aspects, such as probabilities of producing parts outside arbitrary 20 or 3 limits. and pay litle attention tothe practical aspects discussed above. Typical design rules for mechanical components in critical stress application ‘conditions, sch as aicraft and civil engineering stractsral components, require that there must be a factor of safety (say 2) botween the maximum expected stress land the lowor 3e value of the expected strength. Thie approseh is really quite Discrete Variation s arbitary, and oversimplifis the true nature of variations such as strength and loads, a8 described above. Why, for example, select 37? Ifthe strength of the component were truly normally dstbated, about 0.1 percent of components Wwould be weaker thn the sigma value I few componerts are made and used, the probability of ore failing would be very low. However, many are made ard ted, the probity of a sailure among the lager popolation would increase proportionately. I the components used in avery eral application, such a an Jircraftengin suspension hol this probability might be consered too high obo tolerable, Of course there ae often other factors that must be ceneidered such ‘weigh, coat, and the consequences of alu. The era applied to design ofa Slomestc machine might sensibly be less conservative than for @ commercial alrcraft application. _—Theso~ale signa’ approach to achieving high gual ibasedon heideathat, \tany process is convrelied in such a way that ony operations tha exceed plus oF ‘minus 6 of the underlying distribution wil be unaceptabe, hen fewer than one per milicn operations wil fail. The exact quantry ts based on arbitrary ard generally urvealistic assumptions cbout the distribution functions, as described above. (The six ima’ approoc ents other features, such a the we of Wide ‘ange of statistical and othermethodsto identity and reduce variations ofall kinds, and the traning and deployment of specialist. Iti described in Chapter 15) DISCRETE VARIATION “The binomial distribution ‘The binomial distribution describesa situation in which there areonly two outcomes, ‘uch as pase o fail, and the probability remains the same forall tals (Trials which ‘givesuchresulisarecalled Beroullitriels)Thevelore,itisobviously very usefulin QA, ‘nd reality werk The pt fo the Ditomial distibution Is = qiyrt en aaa Tay bovine (2) ‘Thisisthe probability of obtaining x good items and (n — x)bad items ina sample of items, when the probability of selecting a good item isp and of selecting aad item sq. The mean ofthe binomial distribution (from Fqn 213) is given by, wane 239) and the SD from (Eqn 2.17) = (nna)? 239 ey Reliability Mathematics ‘The binomial distribution can only have values at points where x isan integer. The ‘caL.f of the binomial distribution (Le, the probability of obtaning r or fewer suc ‘cesses in tral) is given by m=E (Cre aa Feample28 {A frequent application of the cumulative binomial distribution is in QC acceptance ‘sampling, For example, ifthe acceptance criterion for aproduction lneisthat notmore than t detectives may be found in asample of 2, we can determine the probebiity of acceptance of lot ifthe production process yields 10 percent defectives. From Eqn (20), ro)= 35 Jaros 1957 ‘When approximete values will suffice the nemogram in Fig. 216 can be used Example 27 ‘An aircraft Unilin gear Tia’ 4 (yee: Bapeciace eben thal tyee buat cates on ‘average on 1 landing in 1200, Assuming that tyre bursts occur sindependenily of fone another, and that a safelanding can bemade ifnot moce than 2 tyres burst, what {is the prebabilty of an unsafe lancing? fms the numberof tyres and p is the probability of a tyre bursting n=4 1 ia » 10083 9=(-p) = 099017 ‘hy fang te iy ant methine a 12)~ (lowe syomm m7 + (Jon antoo 7 + ({)iee asrio.se 1n' 00 4159 740.008 325 006 9 +0996 670 831 999 99 997 7 Diseret Variation 3% Prob cor owe eocuanns rn P.P(1-C) Figure 216 Nomgram ofthe cumulauve bemia dstrbuton ‘Therefore the probability ofan uncafe landing is 1 = 0.939 999 997 7= 2.310" “The Poiston distribution Ifevents are Poiston-distibuted they occur ata constant serage rate, with only one of the two outeomer countable, eg, tho rumbor of failures ina given time or defects in along of wire: fo) = Sepp) =0,4,2...) en 3 Reliability Mathematics where jis the mean rate of occurrence. The Poisson distribution can also be con- sidered as an extension ofthe binomial distribution, in whichis considered infinite Figure 217 provides a chart fr the cumulative Foissen distribution. For example, fweneed to Know the probability of not more than three failures occurzing ia 1000 of operation ofa system, when the mean rate of failures is 1 per 1000, we ean read the answer from Fig. 217 a5 0.98 (y = 1/1000, x =3) Since the Poisson distribution can represent the limiting case of the binomial lstribution it gives a good approximation to the binomial distribution, when p o are small and ns large. This is useful in sampling work where the proportion of defectives is low (ie, p< 0.) "The Poisson approximation is 1) =e b ‘This approximation allows us to use Poisson tables charts in appropriate cases and sl sieplifies calculations (2a) pps 0 = (np samples ie pian igi 0, the pty ing at pain oS "heb on (28)oomrean? asm ‘hiss rather more tedious to evaluate than the Posson approximation, The Peisson approximation is evaluated as flows: = 2000 x 0.01 isd Legian tk ‘As the snoemal distribution represents a limiting case of the binomial and Poisson distributions, it can be used to provide a good approximation to these distributions. For example, it can be used when 0.1 > p > 039 and nis large. ‘Then ono (roa) Discrete Variation Ca soanqup won =u neces + € (vane) yt 2 Aaron mind wei o1v0_ 9050 ¥0 £0 Anes ‘vena 000) ss Reliability Mathematics Example2s What is the probability of having not more than 20 failures ifn = 100, Using the bison distribu one Pay = 0.9680 Using the s-normal approximation, w=mp=14 (09) 20-14 3a7 Referring to Appendix 1, Psy = 0.9582. ‘As p— 05, the approximation improves, and we can then use it with smaller values of n. Typically, if p = (4, we can use the approximation with n = 50, STATISTICAL CONFIDENCE Earlier in this chapter we mentioned the problem of statistical confidence, s-confi- lence is the exact faction of times the coufidenc inter! will include the true value, if the experiment is wepeated many times. The confidence intrval i the intrval betsecen the upper and lower s-conidence lite. vconfidence intervals are used in ‘making en assertion about a population given data from a sample. Cleerly, the larger the sample the greater will be our intuitive confidence thatthe estimate of the population parameter will be cose to the true value sonfidence and engineer- {ng confidence must rot be confused; s-confidence takes no account of engineering fr process changes which might make sample data unrepreservatve. Derived ‘confidence values must always be interpreted inthe light of enguneering know- ledge, which might serve to increase or decrease our engineering confidence. confidence limits on continuous variables 1 the population value = follows a s:ncemal distribution, it can be shown that the ‘mean, Sf samples drawn irom it are also s-normally distributed, with variance 2/1 (SD = 9/n"®), The SDof the sample means isalso called the staudarderor ofthe estimate, and is denoted 5, If is rot s-rormally distributed, provided that is lange (> 30) will end toa ‘normal distbation. Tithe distribuiion of + is not excessively skewed (ands unim- ‘dal) the s-normal approximation for ¥ at values of ras small as 6 or 7 may be acceptable “These results are derived from the central limit theorem, mentioned om page 36, ‘They sre of great value in deriving confidence limits on population parameters, ‘based on sample data. In reliability work i is not usually necessary to derive exict Staietiel Confidence Py s-confidence limits and therefore the approximate methods described are quite ad- ‘equate Example2.10 ‘A sample of 100 values has a mean of 2756, witha standard deviation of 1.10. Derive 95 per cent s-confidence limits forthe population mean. (Assume that the sample ‘means are snormally distributed.) ‘In this case, the SD of the sample means, or standard error ofthe estimate s 1 mF Geo? n ‘We can refer tothe table ofthe sneemal cf (Appendlis 1) to tain the 95 percent single-sided s- 30) 075 0>p>15) 080 (6 <15) ‘where ff is the Weibull shape parameter. Therefore, in Example 219, since the Weibull # value appropriate to the normal distribution is > 3, the corrected K-S ccitical value is 0358 » 070 — 0297. Least squares test ‘The least squares goodness-offit testis used to measure the lisa? coralation of data with the equation ofthe straight line which best fits the ploted data. The line vhich ‘most nearly fits 9 st of ploited data is called the regres Iie, and the goodness of fit as derived by the least squares method is called Ue corcaton cogent. “The coreation coeficient is Sa Ey where (Ge. the sample variances) (246) mon te sins am x, yar the coordinate ofthe plied dats, The rogesion in ofthe plot data is Biren by yen 2 fr is positive, it indicates thatthe data are postively corvelated, ie the regression line slope is positive, and vice versa; if | = 1 we have perfect correlation, with all plotted points lying on the regression line; ifr =0 the variates are not linearly Correlated 7 is often used instead of + to indicate correlation, since it provides a more sensitive indication. particularly with probability plots (see Chapter 3). It called the sample coofcen! of determination. inate regression analysis can be used for data that are not lineal correlated if the axes are transformed to inearize the oquation Therefore, the method can be sed for ‘craluating the goodness of fit of data plotted on probability papers (Chapter 3) Sevot of Eoonte a SERIES OF EVENTS (POINT PROCESSES) Situations in which discrete events cccur randomly in 2 continuum (eg, time) ‘cannot be truly represented by a single continuous distribution function. Failures sccurring in repairable systems, aircraft accidents and vehide treffic low past 2 point are examples of series of diserete events. These situations are called stochastic Joint proces. They can be analysed using the statistics of me! series ‘The Poisson distribution function (Eqn 2.41) describes the situation in which vents oecar randomly and at a constant average rate. This situation is described by a homogeneous Poisson process (HPP), An HPP is a stationary point process, since the distibution ofthe number of events in an interval of fixed length does not vary, less of when (where) the interval is sampled “he Poisson distribution function is (roan Eqn 241) ow fj exp( 28) (orn ee) a9, ‘where is the mean rate of occurrence, so that Ais the expected numberof events in (0,2). Figure 28 can be used for estimating the probability of a given number of ‘evens ina given interval Ina nan-homogeneous Poisson process (NHPP) the point process is nor-stationary, sothat the distrbution of the number of events inan interval of fixed length changes as increases. Typically, the discrete events (eg failures) might occur atan increas: ing or decreasing rate "Note that an essential condition of any homogencous Poisson process is thatthe probabilities of events occurring in any period are independent of what has oc furred in preceding periods. An HPP describes a sequence of independently and ‘lentically exponentially distributed (WED) candom variables. A NHEP describes a Sequence of random variables which is nether indeperslenty nor identically di ‘eibuted, Trend analysis ‘hen analysing data from a stochastk: point process itis important to determine Whether the process has a trend, Le. know whether 2 failure rate is increasing, decreasing of constant. We can test for trends by analysing the arroa elses of the event series. The arrival values 21, x2 ..., % are the values of the independent variables (e.g. time) from x= 0 at which each event occurs. Te iatearial ous Xi Nay sony Xe are the intervals between successive events 1, 2, -.., m from ‘= 0. Figure 219 shows the distinction between arrival and intrarrival values. If. isthe peried of observation, then the test statistic for trend is Bu/n—s/2 xo 7B) This is called the centro test oF the Laplace test t compares the centroid of the observed arrival values with the mid-point ofthe pericd of observation, If = 0 250) n Reliability Mathematics ae eater there is no tend, Le. the process is stationary. If U <0 the trond is decreasing, ie the intorrrval values are tending to become larger. Conversely, when U> the trend is increasing, ie. interrsval values are fending to beceme progressively smaller. I the period of observation ends atan event, use (1 ~ 1) instead of and exclude the time to the last event from the summation 2, ‘We can wst the null hypothesis that there is no trend in the chronologically ordered data by testing the value of U against the values of the standard normal variate, z For example, using Appendix 1, if U= 165, for 2=1.85, 442) =035, ‘Therefore we can reject the null hypothesis at the 5 per cent s-signficance level. ‘The centroid tests theoretically adequate if > 4, when the observation interval ends with an event, and ifn > 3, when the interval s terminated at a predetermined time Example 2.20 Arrival values (x) and interarrival values (%)) between 12 successive failures of component areas follows (abservation ends atthe last failure) me 1S sis la % oO a 8 ue ~ ® a3 mn ™m™ a) mo a6 mst Practical Conclusions ” Yay = 5514 (excluding 751) = 5013 a7 Bags 5013-3855 ~ 1.34 Refering to Append 1, for : ~ 154, 2) ~ 094 ai Vax Sa eeee a “Therefor wo can rjc hel hypothesis hat there inno trnd atthe 6 pr ent ignicance level. The interarrval dns are becoming shore. the fae rte is “The existence of «ter inthe dat, es in Example 2.20, ict that he inter aval values ar rot tndepecenty and ely dstibied (ID). Tis a very iimporane point 1 conser inthe analysis of fare dat, a8 wl be explained It Chapter 2 Superimpoted processes fa number of separate stochastic point process combine to form an overall pro- cess, eg, failure processes of individual compenents (or sockets) in a system, these are called superimposed processes. If the individual random variables are ID ‘exponential then the overall process variables also TD exponential and the process is HPP. If the individual variables are ID non-exponontial, the overall process will tend to an HPP. Such + process is callod ronnos! proces. Figure 2.20 shows these processes, (COMPUTER SOFTWARE FOR STATISTICS ‘Computer software is available which can be used to carry out the analytical tech- niques described in this chapter, and in later chapters which describe particular applications. Reference 24 is an excellent program for basic statistics training {and application. Seealso the books homepage (Preface tothe Fourth Aiton. px PRACTICAL CONCLUSIONS Whilst the mathematical methods described in this chapter can be useful for pro- ‘ding inoights and for forecasting, itis smportant tht Unis limitations are apres ated. They are mathematical models, nd they do not necessarily reflect reality the ‘way that deterministic, physics-based formulae do. The importane points that must be borne in mind when applying statstcal methods to engineering are ” Reliability Mathematics [earnest Agere 22) Kote of ccurrenc or superimpose proses Real variation is seldom t-te The most important variation is usually that in the tails, where there is inevitably less (or no) data, the data are more uncertain, and where conventional statistical els can be most misleading, iation can change over time, so thatthe patterns measured atone time might ‘ot represent the true situation at another. We will cover this aspect in more detal fn later chapters. ‘There might be interaction effects between variables, causing combined effects that are more significant than those of indivtual variations. This aspect will be covered in the later chapters. —Variation in engineering is usually made cr influenced by people. Feople do not behave in accordance with any credible mathematical movels Most engineering education in statistics covers only the mathematics, and few statisticians understand the practical aspects of the engineering problems they help to solve. This leads to inappropriate analyses and conclusions. and to a istrast of statistical methods among engineers. We must strike an appropriate balance between using deterministic and statistical ‘methods. For example, if we conduct 2 test in which an item is released from a Ineight, and if it drops the testis a sucess, with results a follows: items tested (Ofailures): 0.110 20 then we could ner thatthe 80 percent s-confidence thatthe reliability is atleast 0.9 would be 0 090 038 099 ‘Thisasoumes tha the daia are entirely statistical, i. we have no prior knowledge of the physics or engineering. On the other hand, if we are confident that weheve such, Biblogrepty * knowledge (in this case, that the force of gravity will aways act on the released items), then we will have 100 per cent confidence in 100 per cent reliability, even. without performing any tests. In such deterministic cases statistical tests and inter- pretations are inappropriate. “Howeever, in many engineering situations can range from deterministic t statis- tical For example, there might be casos when the release mechanism fails to open properly. The causes and effects of variations are often uncertain (particularly when Interactions exist, 29 we must make the best use of our krovledge and use the best ‘methods to explore the uncertaintiss. ‘Statistical testscan, by themselves, generate misleading reults. Wehave discussed this in the cortext of extrapolations beyond the range of measured data. Another ‘example migh be aseries of tests that indicate that an tem operated at high stress, ‘more reliable than when operated at low stress. The results might be due to the items ‘nthe high-sress tests being manufactured using a beter process, othe high stress ‘may actually improve the reliability (eg. a higher temperature might improve the ‘performance ofa seal), othe results might be due to chance and unrepresentative of future tests. The cause ofthe observed reliability difference must be ascertained and, ‘understood in engineering terms. Sometimes ths can be dificult. Uttimately, all of our understanding should be based upon real nowiedge (scientific, human, etc). The statistical methods can provide clues to help us to grin such knowledge. The quotation fram Kendall and Stuart on the inside of the front cover should he the motto forall statistics applications BIBLIOGRAPHY Helpful introductory sources 1. MJ. Moroney Ft fom Figur. Penguin (1985) 2K. Langley, Pica Starisacs Stply Eepare, 20d Edn. Pan (197). 3 WA. Showhai, The Baran: Conta of Menacura Prous, Van Notrourd (1991), IW. W, Hines and D.C: Montgomery, Prot ond Static in Engsering nd Manage Ine Science nl El), Wey (1980 ann Ald Enginering Stato Barres & Noble (197, hati Stats fr Tecimalgy, ind Ean. Chapman te all (197) Ryan, Satistial Method for QualtyImorcoenerd | Wiley (198). ‘Wheeler, Undetanng Veriton: The ey to Managing Chis. SPC Pres (1933, CC Monigomery, GC: Ranger and M: Hubele, Exginering Statist (2nd Elton) Wi On Oe F ». 1 Conand A.W. Lens, The Static Anais of Seri of Eres, Chaprsan & Hall (0948. 11, WI Conover Pract! Now paemutric Satie. Whey (970, ‘More advanced works 12.4.1 Duna, Quality Conran Industral Stites, 5th Ede. fren, Hameed Hino (986) 1B. ALHL Bowker and G J. Lieberman, Engincering Stati, 2nd En, Pence Hall (1972. MG} Hahn andS.5. Shapiro, Statistical Modis n Enginwermg. Wiley (1967) SNL. Johnsen and 5. Kez, Distrbutos Statist 4 Vol Wily (1870, 1. ML. Shooman, ProbbtisicRelibtiry—An Engineering Appreac, McGraw-H (1968). % Reliability Mathematics 17 ES, Pearson and H. O, Harte, Rometrita Tals for Stetson. Biometrika Teast, TE Row and P.Prschan Matin They of Rely Wey (965) EEE Ban ard F. roca See Thor Rly and Tig Halt Richart 20, NM RE Schafer an Sngpwal, Mead Susi nly of ela ind Lip Bate Wiley (1974) oo 7 21. Mende nd ole Nome Seta Meh Whey (07 Eager hn Rady Tr Wie BIL F Mansind A Water yen etttey solr Wey (2). BL W Naan, dplie ets Wy 98) 25 MINAS (len! purpne sates tihng ah pplcaion stare), Mint he 20st Enterprise Dive PTGS, USA, (The MINT Ho Dutbury Pres Bon) ipa erat hacen tbe sat) 26. EW Brey St Neti fo ving, Deon and Mange. Wey tie 2. 6 Ham and Wo Meshes, «Gale rc) Wey (191 Se Sa Sane Epon Uy od Ra doce GOO, ze ‘QUESTIONS 1. Inthe test fring ofa missle, there are some events that are krown to cause the missile to fll reach its target. These events are listed below; together with their approxi- ‘mate probabilies of occurrence during a fight Event Probability (A;} Cloud refection ‘coco (A2) Predpitaton ooUs: (As) Target evasion 0c (AQ) leerenic countermeasures 008 ‘The probabilities of failure if these events occur are: PIA, (03: PUE/As) =0.01; POF/-As) ‘Use Bayes’ theorem (Ean 2.10) to calculate the probability ofeach ofthese eventsbeing ‘the cause inthe event of missile failing reach its target 2. Fora device with a fuilure probability of 0.02 when subjected toa specifi est environ- ment. usethe binemial disribution to calculate the probabilities that ates sample 25 vices will contain (a) no fallures (6) one failure: (€) more than ore flue 13. Repunt question 2 fora failure probability of 02 4 Ropant questions 2 and 3 using the Poisson approximation to the binomial and commenton the avert. 5. Oncof your suppliers hat belatedly eaized that sbout 10 per cnt ofthe batches of ‘Parteular component recenily supplied (0 you hive a manufacturing fault that has Quvetions ” reduced their reliability, There is no extemal or visual means of identifying these subsandard components. Batch identity hus, however, been maintained, $0 your {s to sort batches that have this fault ‘bad! batches) irom the rest good” Batches). An accderated test has been devised such that components from good batches have a failure probablity of 102 whereas these from bad batches have a failure probability of 02. A sampling pln has been devised as follows: (1) Take @ random sample of 25 items from each unknown batch, and subject them to tbe test (2) I there are 0 or I failed components, decide thatthe batch isa ood one (G) Where are two or more failures, decide that the batch is a kad one. “There are risks inthis procedure. In particular, there ae (the risk of deciding that a gow batch is bad: and (i the risk of decding that a bad batch is good. Use Bayes! theorem and your answers to questions 2 and 3 to evaluate thew risks. (@) Explain the circumstances in which you would expect observed faihure mes to canform to an exponential distbution. (0) Faplais the relationship between the exponential and Poisson distibutions ina ‘aliaity content, (6 For equipment with an MTB of 350) calulate the probability of surviving a 200 mission without hilure . A railway tain fited with thve engine/transmieson units that can be assumed to ‘xhibita constant hazard with amean life of 200. na ISh working day, calculate the protabilty of 2 tain having: (a) no failed engine/ transmission unit, (b) not more than one failed unt, (c) not more than two fale units (@) Later Gn Chapter 12) we shall show that failures in complex repairable system fan often be assured 10 result from a Poisson proves. In such o system, 1053H of toting heve been accumaleted, with aires at 994 and 891, Cakulate () the current estimate ofthe system faire rte (@) the eurent estimate ofthe mean time between failures (MBA end (i) dhe lower 99 per cent conience nt forthe NTF (b) Forthe above system, if theresa specification requirement thatthe MTBFshall be atleast 500h, and this must be demonstrated at 90 per cent cunfidence, how much ‘more test running ofthe system, without further failare is required? (6) Explain, using sketches where necessary, the meanings of thefollowing terms used in describing the reliablty behaviour of components and show clearly how they are related toeach other (i) lifetime probability dersity function; i) cumulative distribu tion functor: (i) reiability Furetion: (iv) hazard function. (©) Write down the expression for the cumulative distribution function (ed) (of the two-parameter Weibull distributor. Deine its parameters and_prodoce ‘Sketches 40 show how changing their values influences the caf and the hazard function. 8 Reliability Mathematics 10. Ten components were tested to flue: The ordered ages at faire (hours) were: 7095872; 101.7; 1042; 1062; 111.4; 1125; 1167;143.0, 1503, (@) On the assumption that these times failure are ormally istrbuted, estimate the compenent relabilty and the hazard function () at age 190%; an (i) a age 10h (©) Use a Kolmogorov-Smimov test to see whether itis reasorable to assume normality. 1 A fiywheels retained ona shaft by five bol, which are exch tightened toa specified torgue of 59-£5Nm. A sample of 20 asemblies was checkee for bol torque. The resalls fom the 100 bolts had a mean of 47.2Nm and a standard deviation of 138Ne (4) Assuming that tongues are normally distributed, estimate the proportion below 45Nm. (©) Fora given assembly, what isthe probability of (i) thee being no bolts below 445m il there being atleast one Bolt below 45 Navi there being fewer than ‘neo Boks above 45Na: (it) all five bolts being below 4Nin. (6) nthe overall semple of 10 bos, four were setually four with torques below 445Nn. () Comment on the comparison between this result and your answer to {a above, i) Use this result to obtain a 80 percent rworsided confidence interval for the propertion below 45 Nim. (4) Exphin the meaning ofthe confidence interval in c(i) abore as you would toan inveligent, but non-tecknically-minded, manager. (©) The lowest torque bolt in each assembly was identifed. For these 20 bots, the rnean torque was 455m and the standard deviation 0.48Nm. Assuming an appropriate extreme-vaue dstnbution, calculate the probability that on gen ‘assembly the lowest torque willbe (i) Below 45 Nm; (i) below 441Nm, 12, The following data are the times (hours) between successive fafures ina machining, contre: 46:8; 105,34; 9; 8; 89,138, 75; 186,208; 18,177, Catelate the end statistic (Ean 250) and fest its sigtcance 13, Deserbe four ways in whith the variation of an engineering parameter might be citferent from tht based ypen assuming fat the normal distribution function is tbe correct modal Give an ample in each ese 14, In mos statistical applications the results that matter mos relate othe behaviour of ‘the majority of the population being studied. Why is ths not the case in most ‘engineering applications? Questions 7” 15 Ten items are put on test, until all have fled. The first fake eccurs at 35,000, “operating cycles. Regression aralysis ofthe times to failure shows a good fit to awo- Parmeter Webull distrbution, and the distribution parameters are derived. The specification sates thatthe item should have « By life of 30,000 cycles. Discuss the practical implications assuming thatthe test show's compliance with the specifica ion. 16 In question 15, how might your judgement be infenced if the item is: 8. A tel bolt subject o fatigue loading? A plastic component ina child's toy? A lighting uni? 4. A bearing in a gearbox? ©. A light-emitting diode? (You might like t wy this question after studying Chapters 8 and 9). 17, Explain what s meant by ‘statistical contidenc’. How might statistical confidence “derived from sn experiment be modified inthe light of engineering knowledge? 18 In Example 27, describe three factors that could invalidate the assumption that tyre fallares cocar independent of one another. In what ways might the assumption be more valid for car tyres? 3 Probability Plotting INTRODUCTION tis frequently useful in celabilty engineering © determine which distribution best fits 2 set of data and to derive interval estimates ofthe distrbutien parameters. The ‘mathematical basis for the approaches to these problems Was covered in Chapter 2. Graphical estimation methods can greatly ease this task, and probability plotting ‘papers have been developed for this purpose. These are based upon the cumulative distribution function (-4.f) Fix) ofthe distribution concemed. The axes of probabil- ity ploting papers are transformed in such a way thatthe true cf. plots as 3 ‘straight ine. Therefore ifthe plotted data can be fitted by a straight line, the data fit ‘bution. Further constructions permit the distibution param: Inovledge ofthe statistical mathematics being necessary. This facilitates analysis ‘and presentation of data “The methods described in this chapter can be used to analyse any appropriate data, such as dimensional or parameter measurements, However, their use for ‘analysing reliebiltytime-t-failure data will be emphasized, "Note that probability plotting methods to derive time-tofaihure distribution par~ ‘ameters are only applicable when the data are independently and identically dis- tmibuted (ID). This s usually the case for non-repairable components and systems bat may no! be the case with falure data from repaireble systems. Kelabity data ‘analysis for repairable systems will be covered in Chapter 12 RANKING OF DATA, ‘Mean ranking (normal probability plots) Probability graph papers are based upon plats of the variable of interest against the ‘cumulative percentage probability. The dats therefore need to be ordered and the ‘cumulative probability caleulated. For example, consider the data on times of failure ‘of 12 items (Table 3:1) For the first failure, the camulasive proportion is 1/120r83, per ceat. For the second, the cumulative proportion is 2/12 or 167 per cent, and 90 Ranking of Data a ‘Table 31_Data on testo are of 2 items Cumulaive Mean ak Nedan Timely percent percent rankper Onderne, fature yy edt) edt) cert edt) i m2 s9a/i 77 56 2 wL 67/1) BA BS 3 0 3506/2) BL 217 4 wai 3334/1). ws 28 5 M6 476/12) 385 379 6 M7 5006/12) We? 9 7 7583/12) 38 340 5 al 6676/12) IS a1 8 137 7500/12) 692 702 10 338833071769 783 n 163 S711) a 2 169 son0a/i) 923 3a fn to 12/12 oF 100 per cent for the twelfth failure. However, for probability ploting, Its bette: to make an adjustment 19 allow for the fact that each failure represents @ point on a distribution. Thus considering the whole population of which the 12 Items represent a sample, the times by which 83, 167, .., 100 per cent will have failed in several samples of 12 wil all be randomly distributed. However, the datain “Table 3.1 show a bias in that the first failure is shown much Further from the zero ‘cumulative percentage point than isthe last from 100 per cent in fact it coincides), To overcome this, and thus to improve the accuracy of the estimation, mean or median ranking of cumulative percentages is used for probability plotting, The usual method for mean ranking isto use (n+ 1) in the denominator. instead ‘fn, when calculating the cumulative percentage position. Thus in Table 3.1 the ‘curative percentages (mean ranks) would be 77 per cent 154 per cent 23 per cent ‘Those data are shown ploted on norwal probability paper in Figure 3.1 (irles) The plotied points show a remonably close fit to the stimight line drawn "by eye ‘Therefore, we can say that the date appear to ft the cumilative enormal distbu ton represented by the line. Probability Potting od Amegoad ruuus ee amg ul dveonr arco Chavet se Probability lating Tecuigues ry Median ranking “Mean ranking is the appropriate method fora symmetrical distribution, such asthe s-normal. However, for skewed distributions median renking provides a better ‘correction. The most common approximation used for median ranking is that due to Benard. The ith rank value is given by 3 "Fd ‘where iis the th order value and sis the sample size Median ranking is the method most frequently used in probability ploting, particularly if the data are known not t be normally distributed. Appendix 6 {ives tables of median ranks for samples up to 50. Figure 3.1 shows the data of Table 3.1 plotied using median ranks (triangles). The effect isto alter the slope sightly, but the linearity isnot affected, PROBABILITY PLOTTING TECHNIQUES iting the line Having ranked and plotted the dat, the question that often arises is What i the best stright linet othe daa? (assuming, ofcourse, that there isa reasonable straight line fi). There can be a certain amount of subjectivity or even a temptation to adjust the line a little to fit a preconception. Two basic considerations apply: 1. Since the ploted data are cumulative, the points atthe high cumulative propor- tion end ofthe plot are mere important than the early points. 2. I the conchisions to be derived can be influenced By subjective differences in drawing the line through the data then either there are insufficient data or Shere isnot a good enough fit to the assumed distribution. (The same argument opplies tw choices of which ranking method should be used.) ‘Normally, alin which gives 2 good ‘eyeball it tothe plotted data i satisfactory, ‘and more refined methods wil give results which differ by amounts wel within the 5 per cent significance level. However, a simple and accurate procedure to se, if "ather more objectivity is desired. i to place a transparent rule on the lst point and. ‘draw a line through this peint such that an equal number of point lie elther side of the line Statistical goodness-of-fit tests can be applied, as doscribed in Chaptor 2. It is important to realize that cumulative probability plots are fo a large extent self aligning, since sucseeding points can only continue upwards and to the right Goodies of it tots will nesrly always indicate good correlation with any straight line draven through such points, Therefore itis sometimes prudent to test dhe data against the assured distribution using the methods described. Engineering Know Iralge can abo be of value in determining the likely best distribution. “ Probability Potting ‘confidence limits on plotted data s-confidence limits can be evaluated for the rank values of sample data, as shown in Chapter 2 Appendix 6 includes 5 per cent and 95 per cent median ranks, which can be umed to plot the upper and lower 95 per cent confidence limits on the ed ‘These are shown plotted for the data of Table 31 in Fig. 3.1. ‘Censored data ‘When analysing field or es failure data, we may have to consider samples in which ‘not alitems have failed. A proportion of items may have survived tothe en ofthe test-or some may have been withdravyn prior to the end of the test- Data on such a sample are called censorad data. ems which have survived or which have been removed are called suspended items. The denvation of median ranks for censored data is carried out as ‘olows: 1. List order number (of failed items. 2 List increasing ordered sequence of life values () of failed items. 3. Agint each ae item Ist he number fins whic ave survived na tine and the fist bbetveen that ofthe previous failure and ths failure (or between failure, 44 For each file ito, calewlate the mete order mumber j, sing te Formula fnins ey where eo eee @2) TF (= number of preceding fers) in which mis sample size. 5, Calculate median rank for each failed item, using the formula 7 AS percent es) sa Example 3.1 (Of asample of 30 items, the running times (h fr the failed and surviving items are as shown in Table 3.2 (12 failures. 38 survivors). From Fans (3.1) (33), S0+1-0 T=) ™ = 108 iy = ip +N = 041.08 = Lb Proailty lating Techsiques as ‘Table 32 Running times or aed and surviving ems fora sample of $0 tems uaning Reraing Rening tort me Oem’ tame) emt a os Wm S16 2 ssh SHS A Bk SG R 7m S31 SG s an) 1 lee st mB sa SSS 5S ® 88 is SH 16s 5 S se oi sen 56 o ss &% mM iH 5 % SBS a m 89 oe SM eh 5 me SH as fe mS SS 8 m FR 3 m AOSD ® mm SS Ss % Me Seaman Fok 5051-106 Mee ‘These data (Table 23) are shown plotted on Weibull probability paper in Fig, 3.4 “Sudden death’ testing “Sudden death’ esting is 6 useful technique for obzaining test data quickly on a relatively large sample. The sample is divided inp a number of equal groups % Probability Potting “Table 33. Desvation of median eh for ample 1 Ordered —_-Noof Failure Mfevalues, intervening Mean order, Median rank corderno, tik) survivors Sit) hi 1 percent 1 u 2 im 16 2 ” 0 208 3 3 s 3 3 574 4 %6 1 433 79 3 os U 350 vos ° us ° 00 i202 7 iid 2 O89 07 s w . a7 lo. 3 13 ‘ wiz sa? 0 ist * 1215, Bal an 13. » tear meas 2 20 4 2510 2 (groups, each of A items) by random selection. All tems are then tested unt ene fails in each group. These failures are then used as the data points om the life pet. However, the line obtained represents the distribution oftimes tothe frst failure, in ‘a sample in which all the non failed items are suspended. The times to failure ofthe failed items are renked and plotted, giving sudden death line. For example, Fig. 32 showsa plot ofthe times o first failure often samples of sixeach, withthe rank values based on (10). The fist allure in a sample of six hasa median rank of 101 percent (Appendix 6) Toderive the population median line, tis necessary to drop a vertical from the point where the sudden death line intersects the 50 percent failed ordinate, Aline drawn through the pein where this intersects the 10.91 per cent ordinate, parallel tothe sudden death in, will represent the popsalatin life disteibation. ‘cunt petestage ate WalallProbeility Pots ” Mixed distibations A plot of failure data may relate to one failure mode or to multiple failure modes within an item. Ifa straight line does not fit the failure data, particularly if an ‘obvious change of slope is apparent, the causes of failure shoukl be Investigated to ensure that not mare than one failure mode is included. For example, ater a certain length of time on test, a second failure mode may become apparent, or an item may have two superimposed failure modes. In such cases each failure mode ‘must be isolated and analysed separately, after the data on each have been appro- priately censored (ie. each A-type falure isa survival as far as Btype failures are Cconcemed, and vice versa). However, such separation is appropriate only if the failure procestes are sindependent, ie. there is no interaction. Screened samples {In QC.and reliability work we often deal with samples which have been screened in some way. For example, machined parts will have been inspected, oversize and lndersize parts removed, and electronic parss may have been screeried so that no zerorife (Gead-ce-artval’) parts exis. Screening can show up on protability plots, asacurvature in the tails For example, plotof ime to failure of «fatigue specimen ‘will normally be curved since quailty control will ave removed items of very low strength. In othe: words, there will bea positive minimum life. We will cover this aspect in more detail inthe section on Weibull plotting o lie data. [Estimation of parameters We can also estimate the distribution parameters from the plo. In the example of the normal probability plot given ahove (Fig. 31) tho mean life can be seen 0 be 14h. The standard deviation can be derived by measuring the life interval covered by 34 per cent cither side ofthe 50 per cent point on the probability scale In this cave it L55h, The 95 per cent confidence limits om the mean are 14h and 157h. The estimation of the parameters of other distbutions will be covered in later sections of this chapter, dealing with the Weal and exteeme value distibu- tions. LOGNORMAL PROBABILITY PLOTS: {Lagnormal probability papers sed inexact the sime way a5 normal pecbabity paper except hat the ornate scales constnctd lgarticaly WEIBULL PROBABILITY PLOTS. {In relicblty engineering Weibull probability analysis is the most widely used of the techniques described, due 1 the flexbility of the Welbull dstrbution in describing ‘a number of failure patterns. 88 Probability Potting Weibull transformation The axes of Weibull probability paper are derived by performing the transformation nf (4 astming {ct in sO wee he peomce Weta corns Rw) Ro “This isa straight line o the form y—=ax + ‘Thus the Weibull paper is constructed by having 2 log lg reciprecal ordinate (y) scale, representing cumulative probability of failure or failure percentage, ane the abscissa (x) sales a log scale representing te life value. The slope of a straight line potied on this paper will be J, the shape parameter. Examples of Weibull probability paper are shown in Figures 33 and 34, as produced by Technical and Engineering Aids to Management—TEAM (USA) and Chartwell (UK). On the TEAM paper the slope fs measured by constructing line pparllel tthe line of plotted data, through the estimation point, and reading the value of f atthe intersection with the fl scale (16 in the example shower), The ‘value of fis estimated on the Chartwell paper on the f seal by the intersection of the perpendicular from the estimation point to the line of the ploted data. 7s given by the life value tthe intersecticn withthe estimator line (63.2 percent cumulative {atlure). The Chartwell paper aso includes a scale for estimating the mean life ‘This is derived by reading the survival probability from the P scale, where the f estimator line ntesects, and then reading the fe value relativeto the probability on the cumulative percentage failure scale ‘Another Weibull parameter ften used is the B life. This isthe life by which a ‘given percentage ofthe population will hve filed. For example, the By leis the life at which 10 per cent ofthe population will have failed. Its estimated by reading the appropirate life value against the intersection ofthe lie ofthe plotied data with the 10 per cent cumulative failure line. Fr example, the Bi lie mn Fig. 348 110, (a invicates that this isan estimate from the sample data) Probability papers are available covering various decade ranges. For example, the two papers shown in Figs 33 and 34 cover 3 and 2 decade, spectively. the correct decide range is not available, the data may still be plotted by allowing the plot t reappear on the leit hand side of the paper or by attaching more decade ranges to the ight hand side, Weibull Probability Plots sanded Aaranid gay evan Caney TEAM viet prooney(0.01-989)x 2eyselogannme etna ‘gore 34 Chartwell Webul probity paper foesered data—Fsamype 3.1) Failarestre life ‘The example given o0 far all relate to cases where 7180, there is no faire-fee time: Of course this is not necessarily te ease in reliability work. If am item has finite (positive) fallurefreethme under ‘st, eg, a fatigue test specimen, the failure data vill plot as @ curve, seen conver irom above, since the transformation t0 ‘achieve the Weibull scales assumes that te ata ft a two-parameter distribution. Marepnan, salynuennsit aefopckum npasom Weibull Probability Pots o {1s 1 ruins percentage aire hh b Figure 35 Life plot lure time ‘The effect of finite lie ist shift the age-at-faie scale to the let. It possible to have an apparent negative value for, for example, if the items under teat had accumulated unrecorsed operating time before the start ofthe tert. In this case the curve will appear concave from above. In order to transform such data into a straight line for analysis of the Weibull parameters its necessary to transform the Unree-parameter distribution into a two- [parameter one. To achieve this, equally spaced herizontas are drawn as shown in Fig. 35 with the middle line intersecting the mid-point of the ploted data. The corresponding values oft, fz and ty are then read off the life cal. Them. (= bts =H) 18 GR) G=1) ‘The data are then replotted, with the value of 7 50 derived subtracted from each life ‘value. The life parameters catimated from the plot must then hove the vale of 3 ‘added to give the true lie values Discretion must be used in interpreting data that do not plot as a straight line, since the cause of the non-linearity may be due to the existence of mixed distribu tions or because the date do not fit the Weibull distribution. Its quite likely to be ‘due simply to the randomness in the sample. The failure mechanisms must be ‘studied, and engineering judgement used, to ensure thatthe correct interpretations ane made. It isa commen error fo assume that, because a straight line provides a reasonably good fit to the data, that there is no failure-ree life. However, in most cases wearout failure modes do exhibit a finite failure-ree life. Therefore a value for 7 can sometimes be estimated from knowledge of the product and its application. Alternatively, the time to first failure i often a satisfactory estimate of In these cases the procedure described above is not necessary. Generally, daia on several failure modes in a sysiom are likely 40 fit a two-parameter distribation C=O), but single wearcut failure modes are more likely to have positive values of 7. ey Probability Plotting Table 9A Fail date on 15 eet relays tet alle Median rank Cyc to we Cyc tofaitore, Onter no. percent CO) adjunted (0% 1 45 62 21 2 109 90 4a 3 wa 02 61 4 29 mi 80 5 m5 26 8S 6 310 Ma 03 7 BS 47 108 s 30 er 0 9 55 1B Ms wo 030 as 14 a 5 me 63) 2 Pat Bo 8s 3 25 a7 Re « eer 26 Bs 5 955 3 m3 Frample32 Failure data on 15 electrical relays tested to failure, where the eiterion for failure isa predetermined level of contact resistance, are shown in Table 34 The cycles to failure are plotted in Fig. 3.6. The best fit appears to be the curve ‘shown indicating a finite minimum fe. This conclusicn is aceepted as being valid from knowledge of the relay contact behaviour. ‘The geometrical mid-point on the percentage failure sale gives fo = 13 x 10° and fy are 62 108 and 344 x 108, respectively ‘Subtracting 4.1 from each o the values in the third column gives the results n the fourth coluran. When these are plotted in Figure 2.6 the adjusted data canbe fitted by the stright line shown. From this we can read the cther Weibull distribution parameters: = 17, f)= 16% 10° cyces _=confidence limits on the shape parameter Itis important to derive s-confence limits on the shape parameter f, since deci- sions may be based upon this value. Figure 37 gives factors Fy against sample size for different s-confidence levels (99, 95, 90 per cent) on fl The upper and lower ‘confidence limits are then ye gure 36. Rely ile plot (see Exame 3.2) Ba = fe 3 bin Example 33 Derive the upper and lower confidence llnits if =10, f0=1.8 for C = 90 per cent (Coublesded), Copyrighted material Probability Ploting ly i: + | a | 74 at : | e Marepaan, saquueniiih asropckuM npasom Extreme Value Probability Plating 8 From Fig. 37, Fy = 137; therefore, 6137 =219 16 17 ice. we havea 90 per cont sconfidence that 2.19 > > 1.17, EXTREME VALUE PROBABILITY PLOTTING Extreme value statistics are described in Chapter. TEAM extreme value probability paper is shown in Fig. 38. It uses linear scale forthe random variable and a ‘cumulative probability scale. Two auniliary scales give the refuced variate, and the return period. The return perio is probebility that a future cbocrved value will not exceed the previous value. Thus the return period represents the period over which there is ‘50 per cert probability ofthe event accutring. Bath scales are read by drawing the ‘Verteal through the point of intersection ofthe value ofthe variable with the line of the ploted data Example34 “The breaking strength of long wire was tested, using a sample of 15 equal lengths since the stength of a wire can be considered to depend upon the existence of Imperfections, the extreme value distribution of minimum values might be an appropriate fit The results are shown in Table 35 ard plotted in Fig. 38. Since this would be a distrbution of minimum value it will be leftskewed, and the data are therefore arranged in descending order of magnitude. Plotting the data the other way around would generate a convex curve, viewed from above. A plot of an extreme value distribution of maximum values would be made with the data in ascending order. In this case the probability scale represents the cumulative prob- ability that the breaking strength will be greater than the value indicated, s0 that there is 2.95 per cent probabllity thatthe strength of a wire ofthis Tonge will be freater than 46N. If the wire i longer it will bo likely to bo weaker, since the Probability of it containing extreme value imperfections will be higher. ‘The mode jis estimated directly from the plot. In this case fb 9.5 N.S, the measure of vasabilty, can be derived from the expression (see pege 4): 0578 Probability Potting e 8 8 8 RR TEAM texte value pbabaty 190 visions larepnan, 2auKWeKHEIA aeTope! Hasard Plotting ” ‘Table 35. Breaking strengths of 15 samples of wir of equal length Cumaltve prchabiity percent Breaking strength Gredian inka e)) » 45 %6 m9 va BS ba 39 84 510 555 0 os 70 ms 0 oa i : DRURY Bee nenenn SKerHOLAssAsaa ‘The value of 1/0 is sometimes referred to as the Gumbel slope. The return period 1/(1—p) represents the average value of x (eg. numberof times, or time) between recurrences of a greater (or lesser value than represented by the return period. In Example 34 there is 250-50 chance that a length of wire of breaking, strength lese than 4SN will ccur ina batch of 2 lengths. The return period is used in forecasting the Hkalihood of extrome events euch as floods and high winds, but itis not ofton referred to in reliability work. HAZARD PLOTTING Instead of plotting the cumulative proportion of items failed, we can plot the cumalative hazard function, using Razard papers, This technique has particular advantages when dealing with censored data, a will be shown. ‘The theory undedying hazard ploting is as follows. Frem Chapter 2, we remem- ber thatthe hazard rate hs) is piven by a) T-Fa) his) = en vor om ‘The cumulative hazard function 98 Probability Ploting [mows [taae ez) = Int Fa This relationship allows the derivation of cumulative hazard ploting paper. For ample, for the Weibull distribution, wy =F et hwy = Therefore, n= (2)! If His the cumalative hazard value, log 1 = Flog H+ log 9 B “Thus, Weibull hazard paper is & log-log paper. The nominal slope of the plotted data would be 1/f, When H = 1, x=. Figure 39is an example of Weibull hazard ‘Pape: Hazard papers for normal, lognormal, expenential and type I extreme value distributions are also available The hazard plotting procedure is 1. Tabulate the times to failure in order. 2. For each failure calculate the hazard interval 1 AM. = Camber of items remaining after previows lailures/eensoring 3, For each falure, calculate the cumulative hazard function HAH + MH +o + AH 4. Plot the cumulative hazard against lie value on the chosen hazard paper. ‘As with Weibull probability paper, Weibull hazard paper is based upon the two- ‘Perameter Weibull distribution, ond a non-zero value of failureire if will result in Hasard Plotting o a curved cumulative hazard line. A value for can be derived as described on page 84, except that equal length divisions are measured off on the abscissa instead of on the ordinate ifthe geometrical construction method s being used, Hazard plotting should be used in preference to cumulative protablity ploting when dealing with censored data, or when the data include multiple failure moses and we wish to analyse the overall failure distribution, as well as individual failure ‘modes. Thit i, of course, a type of corsoring,) One tabulation may then bo used, rather than a separate tabulation for each failure mode (se Example 35). However, Since statistical software with plodting capabilites can automatically dealwitheensor ing, probability plotting wsing such software iso more dlficult with censored data. Example 35 Table 3.6 gives distance to fallure on a vehicle shock absorber, taken fram fleet records. The data are tabulated and treated as described above. Fl s the failure ‘mode considered. Note that, compared with the method of dealing with censored. data described earlier (page 64), the calclation of plotting positions is much easier. ‘The data are shown ploted in Fg, 39. In this example, for the failure mode FI, =2., #) = 29000km (the life equiva- lent to 100 per cent cumulative hazard), (This is the 632 per cent c.f. on the probability sale) Dstane —-FAHarard FI Cumulative No. xn) percent hazard per cont 168) ome) 2311/38) 268 207 690 3G6) 7820 405) Bre 3oH) size) 1/38) sx 603) 9860 702) 9a 8o1) uso a0) ee wom) nse nes) 11880 ey ae Be) 123001) oa wos) 120 Boy) BB) a7 ne a) 130 ve 13m wa 14000 900 13M) 50 as was) 70) 525 a8 no 1730 Ban 70 10 Probability Potting Tale (orn) Dine FLAHarard Fi Comulatve No. Gen) percent ivan percent Bue 18450 405) 18960 2514) 18980 213) 2702) 83 8 asa) Bao) 0°) St 2) 230 thes 2 Bren) as ss B@ 2490 M4) 2310) 200 750 sa) Zao % 0) ZeOM Ma ms 7G Ze 8) 28100 CHOOSING THE DISTRIBUTION AND ASSESSING THE RESULTS “The types of failure pattern and modes of fale which can be described by the continuous statistical distributions covered in Chapter 2 and in this chapter have been mentioned inthe relevant sections. The distribution which is likely 0 provide the best fit to a set of data is not always readily apparent, and it can be useful to fellow guidelines which should lead to the most revealing presentation. ‘The steps to be taken are: 1. Consider the situation. What do we expect the failure patter to be? For example, do we expect an increasing or decreasing hazard rate, ora failure-fee lie? >. Plot the ranked data on the appropriate chatt (probability or hazard). 3. Draw the line of best fit afer estimating the failure-iee life if appropriate. If there is good fit to a reasonably large sample (> 10 data points), estimate the [parameters and determine the 5 per cent and 95 per cent confidence limits. 5, Ithe estimated value of fis near 35, try normal paper. ff isbetween 25and 35, tay lognormal paper, but only ifthe sample size is large (> 20) 6. Ifa geod Webull fit cannot be cbtined, and the conditions make 2 fit to an ‘extreme value type I or type Il distribution likly, try extreme value paper: IF ‘extreme value ie unlikely, lock for mulipe failure modes. Slatstical goodnestof-fit tests should be applied to test the fit to the assumed underlying distributions, when finer judgements are to be made and when thore is, 4 Tange database. The refinement inhoront in such touts is not of much valve for ‘Sample sizes less than aout 20, Choosing th Distribution and Assessing the Results one ont ons wal Copyrighted material a Probability Potting fa constant hazard rate (= 1) is apparent thiscan be an indication that multiple failure modes exist or thatthe time-tc-failure data are suspect Ths s often the case ‘with systems in which different parts have diferent ages and individual part ‘operating times are not available. A constant hazard rate can also indicate that {atlures are due to external events, such as faulty use or maintenance. ‘Whilst wearout failure modes are characterized by f > 1, situations can arse in ‘which wearout failures eccur afer a finite falure-free ime, and a value of < Tis ‘obtained. This can happen when only a proportion of the sample is defective, the defect leading to failures after 2 finite time. The Weibull parameters of the ‘wearout failure mode can be derived if ts possible to identify the defective items land analyse their life and failure data separately from that of the non-defective tis very important thatthe physical nature of failures be investigated as pat of ‘any evaluation of failure deta, ifthe results are w be use as. basis for engineering, ‘ction, as opposed to logistic action. For example fan assembly has several failure ‘modes, but the failure data are being used 1 determine requirements for spares and repairs, then we donot need to investigate each mode and the parameters applicable to the mixed distributions are suffice. If, however, we wish to understand the failure modes in order to make improvernents, we must investigate al failures and analyse the distributions separately, since two or more failure modes may have ‘markedly different parameters, yet generate overall failure data which are fitted by a distribution which is diferent to that of any of the underiving ones. Itis also important to ensure thatthe Hime axis chosen is relevant tothe problem, ‘or misleading, results can be generated. For enampie if « numberof items are tested, and the running times recorded. the failure data could show different trends depending upon whether all timer to failure are taken ax cumulative times from, ‘hen the tet cm the first item started, orf individual times to failure are analysed [the items start thei teste at different elapsed or calendar times the results can also bbe misleading if not carefully handled. For example, a trend might be caused by ‘exposure tochanging weathe! conditions, in which case an analysis bases solely on running time could conceal this information. The methods of exploratory data analyss, described in Chapter 12, can be applied when appropriate Finally, its essential that practical engineering crtera are applied to the analysis and toinerpreting the results For example, the earliest time(s) 0 failure are clearly ‘more important than later times; in most cases, so early failures should never be considered as “outliers” unless there is adequate engineering justification for doing so. PROBABILITY PLOTTING FOR BINOMIAL DATA Wome such as alarm systems, standby power supply systems and missile compon- tents can often be most convenient considered a either fancioning or not when ‘needed, and the liability data will thenbe binomially distributed. In those eases we ‘re not concerned with time (or cycles, etc) #9 failure, but only withthe prebabilty (oF failure at the instant consicered, Examples in reability using the binomial stribution were covered ia Chapter 2. Probability Plotting for Binomial Daia ss ‘Tear som eng pow ret 0 ie a a beecess 5 Figure 10 TEAM binomial poting reper (se also Etamp 36) Binomial probability plotting paper can be used for quickly comparing sets of data o for evaluating data against a standard. Figure 310 shows an example of binomial probability paper. I s constructed by having square root scales for the crdinate and the abscissa, so that points Iying cn a straight line through the origin have the same standard deviation. ‘To representa binomial proportion onthe paper. the binomit spit icalculated and plotted. and this is connected to the origin. Thebinomial split isp, (x —p),[ore/(" ~c) {nthe rotation used onthe TEAM paper, where pl iethe quantity with one attribute (eg, defectives) and mis the sample siz, For example, a propertion of 10 per cent ‘wold be plotied at 10 on the c scale and 60 on the racale (not 100), This line then represents all simples with this proportion. »-configence inits canbe drasvn parallel tothe datum line through the appropriate points on the s-confidence scale, 14 Probebilty Potting “Table Reale of tatoon four sample ral Sampleszew Fallurs.c nc 2 5 2 0 2 8 20 5 5 5s wo 6 Binomial papers with different scale ranges are availabe. Since the scales are square rot, they may both be multiplied by multiples of 100to provide factors of 10, inthe variables. Example 36 ‘An item is required to be 90 per cent reliable. Four samples are tested and give results as shown in Table 37, Comment on the results, ‘Since not more than 10 per cent must be defective, the datum point is derived from the binomial split 10/(100 ~ 10) = 10/90. The resuits are then plotted as shown, in Fig. 310. ‘We can now use the confidence limit scale to interpret these results, by drawing, lines parallel tothe requirement line through the scanfidence value require. In this ‘cose wo can see that trial 2 gives the required proportion, However, the res of tial ‘indicates with higher than 90 por cont scontidence thet ii not compliant with the requirement and trials3 and 4 are non-compliant atthe 97-5 percent confidence level ‘Binomial prebebility paper with an extended » sale can be weed as a convenient ‘method of monitoring fest dats for items with a lowe proportion failing, such as ‘lectronic components, a shown in Example 37. Example 3.7 ‘Two types of custom hybrid integrated circuit are used in an instrument. It is ‘required that not more than 0.92 per cent of either type fail during production testing, Test date are shown in Table 38.1 is edded to the cumulative number of {ilures, Pecnuee when plotting binomial date with « soall value ofp, this gives @ Detter estimate of the population proportion j. The failure data are shown plotted in Fig. 3.11. The plo shows that atthe 97.5 per cent confidence level, the type l creat ‘soutsie the specification. The tend shows a statistically signiticant problem in lots 16 to 1, The type 2circut also hasa higher failure ratio than speciied, a the WO per cent confidence level Binomial probability paper as produced by Chartwellis shown in Fig. 3.12. Sample ‘data are plotted in the same way as on the TEAM paper. The cizcular scale gives the percentage defective, as (1 — sino), where ais the value on the degrees scale. This value is read on the vertical scale as shown. For example, Fig, 3.12 shows a plotted Ibincenial split of 10150. This intersects at 757°, giving the percentage defective as (1 sin 757°) = 6.1 per cent. The arcsine transform is used #9 determine scone ‘dence limits, in conjunction with the standard error scale st the top ofthe chart, For Probability Plotting for inom ‘Table 38 Tot dat on to iypesof cus hybrid integrated circu ee Example 37) Data 1s Canes Nofiled Cum Nefaled Com Lots. Noteted ted peliey atl bpeer ait 4 us 18 1 2 ° 1 2 rd 26 ° 2 ' 2 3 138 36s 1 3 0 2 i 1287 381 o 3 0 2 5 463 ald ° 3 1 3 é 1510 728 1 4 0 3 7 i sais 1 5 2 5 5 1384 10399 o 5 0 5 5 5 12168 0 5 0 5 0 188 13817 ° 5 a 5 un 1585 15403 2 7 1 6 a 1383 16756 1 5 a 6 B 1a 18178 o 5 0 5 4 18 19516 o 5 2 5 5 1a 21042 7 5 a 5 16 1 ze 2 0 ° 5 " nz sis 2 B o 5 1 sat 2609 o 6 1 ° 0 6 0 $ at Ea ‘example, the approximate 25 per cent (single-sided) s-confience limits are derived by constracing a cece about the binomial spit point with a radius of two standaed ‘errors. The arcsine transform is then used in the same way to determine the s- ‘confidence limits in this case 10 per cent and 3 per cent. s-confidence limit lines ccan be drawn as on the TEAM paper by constructing lines parallel tothe process line, tangential to the circle Bradley binomial paper Bradley binomial plotting paper allows bincmial data to be analysed and s-confi- ‘dence limits to be derived graphically. Figure 3.13 shows TEAM Bradley binomial plotting paper. The data are plotted by drawing a straight lin from the sample size ‘scale tothe origin, and the value ofthe observed probability is marked on this line. Figure 3.13 shows the plotted data for a sample of 5, of which 10 are defective, ie ‘OB. The s-confidence limits for the population are derived by drawing a horizontal line to intersect the appropriate s-confiderce level (cr standard deviabon) cae. In this case, the 90 per cent s-confidence limits for the defective proportion of the lation would be 0.18 +0086, Bradley binomial paper allows a correction to he made to take account of finite ‘population (ot) size, using the sample-to-lt ratio scale at the top of the graph. Hin this case the lot size is 168, n/N ~ | ~ 0.33. The correction factor fe the=-

You might also like