2/1/2011

Change Language

MIL-217, Bellcore/Telcordia and Other

Quick Links Softw are Nee d He lp? Search GO

Volume 9, Issue 1

Reliability Edge Home

Reliability Prediction Methods for Electronic Products

In today's competitive elec tronic products market, having higher reliability than c ompetitors is one of the key fac tors for suc c ess. To obtain high produc t reliability, consideration of reliability issues should be integrated from the very beginning of the design phase. This leads to the conc ept of reliability prediction. Historic ally, this term has been used to denote the proc ess of applying mathematical models and c omponent data for the purpose of estimating the field reliability of a system before failure data are available for the system. However, the objective of reliability prediction is not limited to predicting whether reliability goals, suc h as MTBF, c an be reac hed. It c an also be used for: Identifying potential design weaknesses. Evaluating the feasibility of a design. Comparing different designs and life- c yc le c osts. Providing models for system reliability/availability analysis. Establishing goals for reliability tests. Aiding in business dec isions such as budget alloc ation and scheduling. Onc e the prototype of a produc t is available, lab tests c an be utilized to obtain more acc urate reliability predic tions. Ac curate predic tion of the reliability of electronic produc ts requires knowledge of the c omponents, the design, the manufac turing proc ess and the expected operating c onditions. Several different approac hes have been developed to ac hieve the reliability predic tion of elec tronic systems and c omponents. Eac h approac h has its unique advantages and disadvantages. Among these approac hes, three main c ategories are often used within government and industry: empiric al (standards based), physic s of failure and life testing. In this artic le, we will provide an overview of all three approaches. First, we will discuss empiric al predic tion methods, whic h are based on the experiences of engineers and on historic al data. Several standards, such as MIL-HDBK-217, Bellc ore/Telcordia, RDF 2000 and China 299B, are widely used for reliability predic tion of electronic products. Next, we will discuss physic s of failure methods, which are based on root-c ause analysis of failure mec hanisms, failure modes and stresses. T his approach is based upon an understanding of the physic al properties of the materials, operation processes and tec hnologies used in the design. Finally, we will disc uss life testing methods, whic h are used to determine reliability by testing a relatively large number of samples at their spec ified operation stresses or higher stresses and using statistic al models to analyze the data.

Empirical (or Standards Based) Prediction Methods

Empirical predic tion methods are based on models developed from statistical curve fitting of historic al failure data, whic h may have been collected in the field, in-house or from manufac turers. These methods tend to present good estimates of reliability for similar or slightly modified parts. Some parameters in the c urve func tion c an be modified by integrating engineering knowledge. The assumption is made that system or equipment failure c auses are inherently linked to c omponents whose failures are independent of each other. There are many different empiric al methods that have been created for spec ific applic ations. Some have gained popularity within industry in the past three decades. The table below lists some of the available prediction standards and the following sec tions describe three of the most commonly used methods in a bit more detail.

reliasoft.com//prediction_methods.htm

1/11

2/1/2011

MIL-217, Bellcore/Telcordia and Other

MIL-HDBK-217 Predictive Method MIL-HDBK- 217 is very well known in military and c ommerc ial industries. It is probably the most internationally rec ognized empirical predic tion method, by far. The latest version is MIL-HDBK217F, whic h was released in 1991 and had two revisions: Notic e 1 in 1992 and Notice 2 in 1995. The MIL-HDBK-217 predic tive method consists of two parts; one is known as the parts c ount method and the other is c alled the part stress method [1]. The parts c ount method assumes typic al operating c onditions of part c omplexity, ambient temperature, various elec tric al stresses, operation mode and environment (c alled referenc e conditions). The failure rate for a part under the referenc e conditions is calc ulated as:

where: ref is the failure rate under the referenc e c onditions.

i is the number of parts.

Sinc e the parts may not operate under the referenc e c onditions, the real operating c onditions will result in failure rates that are different from those given by the "parts count" method. Therefore, the part stress method requires the specific parts complexity, applic ation stresses, environmental fac tors, etc . (c alled Pi fac tors). For example, MIL-HDBK-217 provides many environmental c onditions (expressed as E) ranging from "ground benign" to "c annon launc h." The standard also provides multi-level quality specific ations (expressed as Q). The failure rate for parts under spec ific operating conditions c an be calc ulated as:

where:

S is the stress factor. T is the temperature fac tor. E is the environment fac tor. Q is the quality factor. A is the adjustment fac tor.
Figure 1 shows an example using the MIL- HDBK-217 method (in ReliaSofts Lambda Predic t software) to predic t the failure rate of a c eramic c apac itor. Ac c ording to the handbook, the failure rate of a commerc ial c eramic c apac itor of 0.00068 mF c apac itanc e with 80% operation voltage, working under 30 degrees ambient temperature and "ground benign" environment is 0.0216/106 hours. The c orresponding MTBF (mean time before failure) or MTTF (mean time to failure) is estimated to be 46,140,368 hours.

reliasoft.com//prediction_methods.htm

2/11

2/1/2011

MIL-217, Bellcore/Telcordia and Other

Figure 1: MIL-HDBK-217 capacitor failure rate example Bellcore/Telcordia Predictive Method Bellc ore was a telec ommunications researc h and development company that provided joint R&D and standards setting for AT&T and its c o- owners. Bec ause of dissatisfaction with military handbook methods for their c ommerc ial produc ts, Bellc ore designed its own reliability prediction standard for commerc ial telec ommunic ation products. In 1997, the c ompany was ac quired by Scienc e Applic ations International Corporation (SAIC) and the c ompany's name was c hanged to Telc ordia. Telc ordia continues to revise and update the standard. The latest two updates are SR- 332 Issue 1 (May 2001) and SR-332 Issue 2 (September 2006), both c alled "Reliability Prediction Proc edure for Elec tronic Equipment." The Bellc ore/Telc ordia standard assumes a serial model for elec tronic parts and it addresses failure rates at the infant mortality stage and at the steady- state stage with Methods I, II and III [2-3]. Method I is similar to the MIL-HDBK- 217F parts c ount and part stress methods. The standard provides the generic failure rates and three part stress factors: devic e quality factor ( Q), elec tric al stress factor ( S) and temperature stress factor (T). Method II is based on combining Method I predic tions with data from laboratory tests performed in acc ordance with specific SR-332 c riteria. Method III is a statistic al predic tion of failure rate based on field trac king data c ollec ted in ac cordance with specific SR-332 c riteria. In Method III, the predic ted failure rate is a weighted average of the generic steady-state failure rate and the field failure rate. Lambda Predict has implemented Methods I and II, and Method III will be added in the next version. Figure 2 shows an example in Lambda Predict using SR- 332 Issue 1 to predic t the failure rate of the same c apacitor in the previous MIL- HDBK-217 example (shown in Figure 1). The failure rate is 9.654 Fits, whic h is 9.654 / 109 hours. In order to c ompare the predicted results from MIL- HBK-217 and Bellc ore SR-332, we must convert the failure rate to the same units. 9.654 Fits is 0.000965 / 106 hours. So the result of 0.0216 / 106 hours in MIL-HDBK- 217 is muc h higher than the result in Bellc ore/Telcordia SR- 332. There are reasons for this variation. First, MIL-HDBK- 217 is a standard used in the military so it is more c onservative than the commerc ial standard. Sec ond, the underlying methods are different and more fac tors that may affect the failure rate are considered in MIL- HDBK-217.

reliasoft.com//prediction_methods.htm

3/11

2/1/2011

MIL-217, Bellcore/Telcordia and Other

Figure 2: Bellcore capacitor failure rate example RDF 2000 Predictive Method RDF 2000 is a reliability data handbook developed by the Frenc h telecommunic ations industry. This standard provides reliability prediction models for a range of electronic c omponents using cyc ling profiles and applic able phases as a basis for failure rate c alc ulations [4]. RDF 2000 provides a unique approach to handle mission profiles in the failure rate predic tion. Component failure is defined in terms of an empiric al expression c ontaining a base failure rate that is multiplied by factors influenc ed by mission profiles. These mission profiles c ontain information about how the c omponent failure rate may be affec ted by operational cyc ling, ambient temperature variation and/or equipment switc h on/off temperature variations. RDF 2000 disregards the wearout period and the infant mortality stage of produc t life based on the assumption that, for most elec tronic c omponents, the wearout period is never reac hed bec ause new produc ts will replac e the old ones before the wearout oc curs. For c omponents whose wearout period is not very far in the future, the normal life period has to be determined. The infant mortality stage failure rate is c aused by a wide range of factors, such as manufac turing proc esses and material weakness, but c an be eliminated by improving the design and production proc esses (e.g. by performing burn-in). As an example, the empiric al expression formula for a ceramic c apac itor of c lass I is given by:

where: ( t)i is the temperature fac tor related to the ith junction temperature of the capac itor mission profile.

i is the working time ratio of the c apac itor for the ith junc tion temperature of the mission
profile. on is the total working time ratio of the c apac itor, with on + off = 1. ( n)i is the ith influenc e fac tor related to the annual cyc les number of thermal variations seen by the package, with the amplitude T.

Ti is the thermal amplitude variation of the ith mission profile.

reliasoft.com//prediction_methods.htm

4/11

2/1/2011

MIL-217, Bellcore/Telcordia and Other

Figure 3 shows the implementation of the failure rate predic tion using RDF 2000 in Lambda Predict.

Figure 3: RDF 2000 capacitor failure rate example Discussion of Empirical Methods Although empirical prediction standards have been used for many years, it is always wise to use them with c aution. The advantages and disadvantages of empiric al methods have been disc ussed a lot in the past three dec ades. A brief summary from the public ations in industry, military and academia is presented next [5- 9]. Advantages of empirical methods: 1. Easy to use, and a lot of c omponent models exist. 2. Relatively good performance as indicators of inherent reliability. 3. Provide an approximation of field failure rates. Disadvantages of empirical methods: 1. A large part of the data used by the traditional models is out- of-date. 2. Failure of the c omponents is not always due to component-intrinsic mec hanisms but c an be c aused by the system design. 3. The reliability predic tion models are based on industry-average values of failure rate, which are neither vendor-specific nor devic e-specific . 4. It is hard to c ollect good quality field and manufac turing data, whic h are needed to define the adjustment fac tors, suc h as the Pi factors in MIL-HDBK-217.

Physics of Failure Methods

In c ontrast to empiric al reliability predic tion methods, whic h are based on the statistic al analysis of historic al failure data, a physic s of failure approac h is based on the understanding of the failure mec hanism and applying the physic s of failure model to the data. Several popularly used models are disc ussed next. Arrhenius's Law One of the earliest and most suc cessful ac c eleration models predic ts how the time-to-failure of a system varies with temperature. This empiric ally based model is known as the Arrhenius equation. Generally, c hemical reac tions c an be ac c elerated by increasing the system temperature. Sinc e it is a chemic al process, the aging of a c apac itor (such as an electrolytic capacitor) is ac c elerated by increasing the operating temperature. The model takes the following form.

where:

reliasoft.com//prediction_methods.htm

5/11

2/1/2011

MIL-217, Bellcore/Telcordia and Other

L(T ) is the life c haracteristic related to temperature. A is the sc aling fac tor. Ea is the activation energy. k is the Boltzmann c onstant. T is the temperature.
Eyring and Other Models While the Arrhenius model emphasizes the dependency of reac tions on temperature, the Eyring model is commonly used for demonstrating the dependenc y of reactions on stress factors other than temperature, suc h as mec hanic al stress, humidity or voltage. The standard equation for the Eyring model [10] is as follows:

where:

L(T ,S ) is the life c harac teristic related to temperature and another stress. A, , B and C are c onstants. S is a stress fac tor other than temperature. T is absolute temperature.
Acc ording to different physic s of failure mec hanisms, one more term (i.e. stress) can be either removed or added to the above standard Eyring model. Several models are similar to the standard Eyring model. They are: Two Temperature/Voltage Model:

Three Stress Model (Temperature-Voltage-Humidity):

Corrosion Model: Elec tronic devic es with aluminum or aluminum alloy with small perc entages of c opper and silicon metallization are subjec t to c orrosion failures and therefore c an be described with the following model [11]:

where:

B0 is an arbitrary sc ale fac tor. is equal to 0.1 to 0.15 per % RH. f(V) is an unknown function of applied voltage, with empiric al value of 0.12 to 0.15.
Hot Carrier Injection Model: Hot carrier injec tion desc ribes the phenomena observed in MOSFETs by whic h the c arrier gains suffic ient energy to be injec ted into the gate oxide, generate interfac e or bulk oxide defects and degrade MOSFETs charac teristic s suc h as threshold voltage, transc onductanc e, etc . [11]: For n- channel devic es, the model is given by:

where:

B is an arbitrary sc ale fac tor.

Isub is the peak substrate c urrent during stressing.

N is equal to a value from 2 to 4, typically 3. Ea is equal to -0.1eV to -0.2eV.

For p- channel devic es, the model is given by:

where:

reliasoft.com//prediction_methods.htm

6/11

2/1/2011

MIL-217, Bellcore/Telcordia and Other

B is an arbitrary sc ale fac tor. Igate is the peak gate c urrent during stressing. M is equal to a value from 2 to 4. Ea is equal to -0.1eV to -0.2eV.
Sinc e elec tronic produc ts usually have a long time period of useful life (i.e. the c onstant line of the bathtub c urve) and c an often be modeled using an exponential distribution, the life charac teristics in the above physic s of failure models c an be replac ed by MTBF (i.e. the life charac teristic in the exponential distribution). However, if you think your produc ts do not exhibit a c onstant failure rate and therefore c annot be desc ribed by an exponential distribution, the life c harac teristic usually will not be the MTBF. For example, for the Weibull distribution, the life charac teristic is the scale parameter eta and for the lognormal distribution, it is the log mean. Black Model for Electromigration Elec tromigration is a failure mec hanism that results from the transfer of momentum from the elec trons, whic h move in the applied electric field, to the ions, whic h make up the lattic e of the interc onnec t material. The most c ommon failure mode is "c onductor open." With the dec reased structure of Integrated Circ uits (ICs), the inc reased c urrent density makes this failure mec hanism very important in IC reliability. At the end of the 1960s, J. R. Blac k developed an empirical model to estimate the MTTF of a wire, taking electromigration into consideration, whic h is now generally known as the Black model. The Blac k model employs external heating and inc reased c urrent density and is given by:

where:

A0 is a c onstant based on the cross- sec tional area of the interc onnec t. J is the c urrent density. J threshold is the threshold c urrent density. E a is the ac tivation energy. K is the Boltzmann constant. T is the temperature. N is a sc aling fac tor.
The c urrent density ( J) and temperature (T) are factors in the design proc ess that affec t elec tromigration. Numerous experiments with different stress conditions have been reported in the literature, where the values have been reported in the range between 2 and 3.3 for N, and 0.5 to 1.1eV for Ea. Usually, the lower the values, the more conservative the estimation. Coffin-Manson Model for Fatigue Fatigue failures c an oc c ur in elec tronic devic es due to temperature c ycling and thermal shoc k. Permanent damage ac cumulates eac h time the devic e experienc es a normal power-up and power- down c ycle. These switc h cyc les c an induc e c yclical stress that tends to weaken materials and may cause several different types of failures, suc h as dielectric /thin- film cracking, lifted bonds, solder fatigue, etc . A model known as the (modified) Coffin-Manson model has been used succ essfully to model c rac k growth in solder due to repeated temperature cyc ling as the devic e is switc hed on and off. This model takes the form [9]:

where:

Nf is the number of cyc les to failure. is a c oefficient. f is the cyc ling frequency. T is the temperature range during a cyc le. is the c yc ling frequenc y exponent. is the temperature exponent. G(Tmax) is equal to:

whic h is an Arrhenius term evaluated at the maximum temperature in eac h c ycle.

reliasoft.com//prediction_methods.htm

7/11

2/1/2011

MIL-217, Bellcore/Telcordia and Other

Three factors are usually considered for testing: maximum temperature (T max), temperature range (T) and cyc ling frequency (f). The ac tivation energy is usually related to c ertain failure mec hanisms and failure modes, and c an be determined by c orrelating thermal c ycling test data and the Coffin- Manson model. Discussion of Physics of Failure Methods A given elec tronic c omponent will have multiple failure modes and the c omponent's failure rate is equal to the sum of the failure rates of all modes (i.e. humidity, voltage, temperature, thermal c yc ling and so on). The system's failure rate is equal to the sum of the failure rates of the c omponents involved. In using the above models, the model parameters can be determined from the design spec ific ations or operating c onditions. If the parameters c annot be determined without c onduc ting a test, the failure data obtained from the test can be used to get the model parameters. Software produc ts such as ReliaSoft's ALTA can help you analyze the failure data. We will give an example of using ALTA to analyze the Arrhenius model. For this example, the life of an electronic component is c onsidered to be affec ted by temperature. The c omponent is tested under temperatures of 406, 416 and 426 Kelvin. The usage temperature level is 400 Kelvin. The Arrhenius model and the Weibull distribution are used to analyze the failure data in ALTA. Figure 4 shows the data and c alc ulated parameters. Figure 5 shows the reliability plot and the estimated B10 life at the usage temperature level.

Figure 4: Data and analysis results in ALTA with the Arrhenius- Weibull model

reliasoft.com//prediction_methods.htm

8/11

2/1/2011

MIL-217, Bellcore/Telcordia and Other

Figure 5: Reliability vs. Time plot and calculated B10 life From Figure 4, we c an see that the estimated ac tivation energy in the Arrhenius model is 0.92. Note that, in ALTA, the Arrhenius model is simplified to a form of:

Using this equation, the parameters B and C c alc ulated by ALTA c an easily be transformed to the parameters desc ribed above for the Arrhenius relationship. Advantages of physics of failure methods: 1. Ac c urate prediction of wearout using known failure mechanisms. 2. Modeling of potential failure mechanisms based on the physic s of failure. 3. During the design process, the variability of eac h design parameter c an be determined. Disadvantages of physics of failure methods: 1. Need detailed c omponent manufac turing information (suc h as material, process and design data). 2. Analysis is complex and c ould be c ostly to apply. 3. It is difficult to assess the entire system.

Life Testing Method

As mentioned above, time-to-failure data from life testing may be inc orporated into some of the empirical prediction standards (i.e. Bellc ore/Telc ordia Method II) and may also be nec essary to estimate the parameters for some of the physic s of failure models. However, in this sec tion of the artic le, we are using the term life testing method to refer specific ally to a third type of approac h for predic ting the reliability of electronic produc ts. With this method, a test is conduc ted on a suffic iently large sample of units operating under normal usage conditions. Times-to-failure are rec orded and then analyzed with an appropriate statistic al distribution in order to estimate reliability metric s suc h as the B10 life. This type of analysis is often referred to as Life Data Analysis or Weibull Analysis. ReliaSoft's Weibull++ software is a tool for conduc ting life data analysis. As an example, suppose that an IC board is tested in the lab and the failure data are rec orded. Figure 6 shows the data entered into Weibull++ and analyzed with the 2- parameter Weibull lifetime distribution while Figure 7 shows the Reliability vs. Time plot and the c alculated B10 life for the analysis.

reliasoft.com//prediction_methods.htm

9/11

2/1/2011

MIL-217, Bellcore/Telcordia and Other

Figure 6: Data and analysis results in Weibull++ with the Weibull distribution

Figure 7: Reliability vs. Time plot and calculated B10 life for the analysis Discussion of the Life Testing Method The life testing method c an provide more information about the produc t than the empiric al prediction standards. T herefore, the predic tion is usually more acc urate, given that enough samples are used in the testing.

reliasoft.com//prediction_methods.htm

10/11

2/1/2011

MIL-217, Bellcore/Telcordia and Other

The life testing method may also be preferred over both the empiric al and physic s of failure methods when it is nec essary to obtain realistic predic tions at the system (rather than component) level. This is bec ause the empiric al and physics of failure methods c alc ulate the system failure rate based on the predic tions for the c omponents (e.g. using the sum of the component failure rates if the system is c onsidered to be a serial c onfiguration). This assumes that there are no interac tion failures between the c omponents but, in reality, due to the design or manufac turing, c omponents are not independent. (For example, if the fan is broken in your laptop, the CPU will fail faster bec ause of the high temperature.) Therefore, in order to c onsider the c omplexity of the entire system, life tests can be c onduc ted at the system level, treating the system as a "blac k box," and the system reliability c an be predic ted based on the obtained failure data.

Conclusions
In this article, we discussed three approac hes for electronic reliability predic tion. The empiric al (or standards based) methods c an be used in the design stage to quic kly obtain a rough estimation of produc t reliability. The physic s of failure and life testing methods c an be used in both design and production stages. In physics of failure approac hes, the model parameters can be determined from design specs or from test data. On the other hand, with the life testing method, sinc e the failure data from your own particular products are obtained, the prediction results usually are more ac curate than those from a general standard or model.

References
[1] MIL-HDBK- 217F, Reliability Predic tion of Electronic Equipment, 1991. Notice 1 (1992) and Notice 2 (1995). [2] SR-332, Issue 1, Reliability Prediction Proc edure for Elec tronic Equipment, Telc ordia, May 2001. [3] SR-332, Issue 2, Reliability Predic tion Proc edure for Electronic Equipment, Telcordia, September 2006. [4] ITEM Software and ReliaSoft Corporation, RS 490 Course Notes: Introduc tion to Standards Based Reliability Predic tion and Lambda Predic t, 2006. [5] B. Fouc her, J. Boullie, B. Meslet and D. Das, "A Review of Reliability Prediction Methods for Elec tronic Devices," Mic roelec tron. Wearout., vol. 42, no. 8, August 2002, pp. 1155- 1162. [6] M. Pec ht, D. Das and A. Ramarkrishnan, "The IEEE Standards on Reliability Program and Reliability Prediction Methods for Elec tronic Equipment," Mic roelec tron. Wearout., vol. 42, 2002, pp. 1259- 1266. [7] M. Talmor and S. Arueti, "Reliability Prediction: The Turnover Point," 1997 Proc . Ann. Reliability and Maintainability Symp., 1997, pp. 254- 262. [8] W. Denson, "The History of Reliability Predic tion," IEEE Trans. On Reliability, vol. 47, no. 3SP, September 1998. [9] D. Hirsc hmann, D. Tissen, S. Schroder and R.W. de Doncker, "Reliability Prediction for Inverters in Hybrid Elec tric al Vehic les," IEEE Trans. on Power Elec tronic s, vol. 22, no. 6, November 2007, pp. 2511-2517. [10] NIST Information Tec hnology Library. [Online doc ument] Available HTTP: www.itl.nist.gov [11] Semic onductor Devic e Reliability Failure Models. [Online doc ument] Available HTTP: www.sematech.org/doc ubase/doc ument/3955axfr.pdf

[Editorial Note: In the printed edition of Volume 9, Issue 1, there were two errors that have been corrected in this online version. We apologize for any inconvenienc e. 1) 9.654 Fits is 9.654 / 109 hours (rather than 1010). 2) In the equations for hot c arrier injection models, "Ea is equal to -0.1eV to -0.2eV."]

C opyright 1992 - 2011 ReliaSoft C orporation. All Rights Reserved Privacy Statement | Terms of Use | C ontact | About Us

reliasoft.com//prediction_methods.htm

11/11