1、_SAE Technical Standards Board Rules provide that: “This report is published by SAE to advance the state of technical and engineering sciences. The use of this report is entirely voluntary, and its applicability and suitability for any particular use, including any patent infringement arising theref
2、rom, is the sole responsibility of the user.”SAE reviews each technical report at least every five years at which time it may be revised, reaffirmed, stabilized, or cancelled. SAE invites your written comments and suggestions.Copyright 2017 SAE InternationalAll rights reserved. No part of this publi
3、cation may be reproduced, stored in a retrieval system or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise, without the prior written permission of SAE.TO PLACE A DOCUMENT ORDER: Tel: 877-606-7323 (inside USA and Canada)Tel: +1 724-776-4970 (out
4、side USA)Fax: 724-776-0790Email: CustomerServicesae.orgSAE WEB ADDRESS: http:/www.sae.orgSAE values your input. To provide feedback on thisTechnical Report, please visithttp:/standards.sae.org/J3083_201703SURFACE VEHICLERECOMMENDED PRACTICEJ3083 MAR2017Issued 2017-03Reliability Prediction for Automo
5、tive Electronics Based on Field Return DataRATIONALEIn early design activities (typically before the hardware is built), a reliability prediction is often required for the electronic components and systems in order to assess their future reliability and in many cases to meet customer specifications.
6、Those specifications may include the allocated reliability for a particular electronic unit and in the cases of functional safety products to meet the ASIL (Automotive Safety and Integrity Level) requirement specified by the international functional safety standard ISO 26262.This Recommended Practic
7、e (RP) document will provide guidance on performing reliability predictions for automotive electronic products utilizing field return data or any other types of failure data available to an automotive electronics supplier. This document will cover the possible sources of data, types of the data requ
8、ired, ways to collect it, and the methodology of how to process these data to calculate the failure rates and the expected reliability. This document will also include a case study based on the data obtained by Delphi Electronics however, there are situations when a system has a redundancy and needs
9、 to be modeled accordingly.4.4 Failure Rate Calculations: Assumptions and LimitationsThe following considerations are important and should be taken into account to avoid misleading reliability prediction results obtained for electronic components and systems.1. It is important to remember that any f
10、orm of reliability prediction, either handbook or field- based, utilizes prior failure history with components and/or electronic units. The assumed similarity between the old and new components is the basis of all these methods, which is also their limiting factor. New technologies and new manufactu
11、ring techniques bring new failure modes, new failure sites and failure mechanisms, e.g., disruptive technologies for thesemiconductors (nano-structures) and/or package assembly (gold/copper wires, solder balls/copper pillars, stacked dies, etc.), and package types (leaded SMD/BGA, outline size, etc.
12、), as well as the assembly on PCB. Any correlation between the already used and new components is only possible, as long as they do not involve substantially different design concepts. When they do, that correlation becomes significantly weaker.2. These methods also assume comparable application mis
13、sion profiles for the automotive electronics, such as passenger compartment, engine compartment, door, permanent power-on in sleeping mode, etc.It is important to remember that after the warranty period expires, the ratio of parts failed in the field to the number of parts returned to the manufactur
14、er for engineering analysis is expected to drop significantly due to the alternative sources of repair. Therefore, in order to extract the failure rates from the post-warranty data, an accurate assessment of that ratio will be required. If the assessment is not possible due to lack of data, it is st
15、rongly recommended to compare the FIT numbers derived from only warranty data with the numbers from all data including out of warranty data. If there is a big variation between these two FIT numbers, the conservative one (i.e., higher failure rate) should be taken. This is especially important for t
16、he safety applications. Many companies in the supply chain do not or cannotperform a careful time and cost spending analysis of returned parts or, as long as the return rate is below an agreed upon low level, have instead a financial compensation agreement with their suppliers and customers. As a re
17、sult, their available data contain many electronic units reported as failed due to EOS-damages of semiconductors because of operation out-of-spec (hot plugging, overvoltage, etc.), and No-Trouble-Found units (unqualified replacement in field) but no data on inherent component failures. The real comp
18、onent failure rate then would be significantly lower. A very comprehensive and extensive analysis from Bosch 16 has proven that the majority of destroyed ICs in returned failed ECUs was caused by operation out-of-spec, like hot-plugging or overload conditions.SAE INTERNATIONAL J3083 MAR2017 Page 10
19、of 223. Failure rate prediction of semiconductor devices present significant challenge to all the averaging methods including handbooks and field-based calculations. Considering the variety of applications and complexities in the todays semiconductor devices ranging from embedded to system-on-chip a
20、nd system-in-package, individual FIT computation would need to consider various elements and characterize thermal-mechanical fluctuations in the substrate-to-die.Traditional FIT that relies heavily on constant voltage and concentrated high temperature locale may prove to be somewhat simplistic, wher
21、eas IC performance can significantly modulate from infant-mortality region to the wear-out area. Also, user application and environment can be critical contributors (see Section 9). In order to correctly capture all these variations and further improve the accuracy of the forecasting method describe
22、d in this document, an approach that focuses on physics of failure (PoF) to target the applications landscape from usage to environmental condition would need to be considered. This approach is commonly referred as Knowledge-Based-Test-Methodologyand has been described in JESD94 30. Once an appropri
23、ate model reflecting physics of failure has been identified, either from historical data, or literature, or empirically developed, then FIT can be derived, and verified using stress tests that forcibly accelerates the devices lifetime by a particular parameter (temperature, humidity, voltage, etc.).
24、However, for practical purposes the averaging methods are still widely utilized due to lack of resources required toclosely adhere to the PoF methodology described in 30.4. Other important considerations and limitations of the reliability prediction have been covered in IEEE 1413 3 and SAE J1938 15.
25、 Despite all the limiting factors discussed above, the prediction error based on the field-based data is expected to be lower than that based on the handbooks, due to the fact that typically the field based data is updated more frequently than the handbooks. Hence, it is more reflective of changes i
26、n the new technology than any of the paper standards.5. CALCULATING FAILURE RATESIn addition to warranty data, many automotive suppliers or OEMs collect field return information and other types of detailed failure data for further engineering analysis purposes. Often, this data contains comprehensiv
27、e repair information on electronic systems as well as individual components. In that case, the data can be used to calculate failure rates for electronic components which can, in turn, be used for system reliability predictions. These types of data sources exist in many companies and are usually man
28、aged by their quality, reliability, and/or customer satisfaction organization.The analysis method presented here is based on the field return data, which can be collected via internal warranty databases and/or electronic unit repair information from various companys field return sources, such as rep
29、air centers, FRACAS, detailed warranty return data, etc. For more details on the possible data sources please see Sections 5.3 and 8.5.1 Simplified CalculationsDuring continuous testing or field data collection, the mean time between failures can be calculated as the total test timeor total time in
30、the field T among all observed units divided by the number of failures.TMTBFk6 (Eq. 12)where:k = total number of failures accumulated by all the units in the population, k 01NiiTT6 total time accumulated by all the N units in the populationTi= total time accumulated by the ithcomponent in the popula
31、tionIn the case of accelerated and/or bench testing 1iNTiiTAFT6 uSAE INTERNATIONAL J3083 MAR2017 Page 11 of 22TTi= total field time accumulated by the ithcomponent on testAFi= acceleration factor for the ithcomponent on testUnder the assumption of the exponential distribution Equation 1, the constan
32、t failure rate O is the inverse of the MTBF and hence can be approximated by the ratio of the total number of failures to the total time accumulated by the componentduring the test or a field operation shown in Equation 12.kTO6 (Eq. 13)Equations 12 and 13 are approximations and do not include any co
33、nfidence intervals on either MTBF or O.5.2 Statistical Confidence Intervals on Failure RatesIn order to assign confidence intervals to failure rates they need to be modeled with a statistical distribution. The $2(chi-squared) distribution has been widely used in quality and reliability engineering a
34、nd, due to its properties is typically used for statistical testing, goodness of fit tests and evaluating confidence. It can be used to put confidence intervals on the number of failures we expect to see during the time interval T6, from Equation 12, when a product has an assumed constant failure ra
35、te. That number of failures according to Equation 12 can be expressed as T6 /MTBF. According to 5, the probability that the number of failures for a given period of time T within the confidence bounds (1-D) will be:22/2,2( 1) 1 /2,2( 1)2Pr 1kkTMTBFDDFFDdd (Eq. 14)where:F2D,2(k+1) = chi-squared distr
36、ibutionk = number of observed failuresD = risk factor = 1 ConfidenceEquation 14 can also be used to calculate confidence bounds on MTBF and consequently on the value of the failure rate O. In the cases of time terminated test or field observations, the upper bound for a failure rate can be estimated
37、 similar to Equation 14 (see for example 4 Chapter 12) as:2,2( 1)2kTDFO6d(Eq. 15)where:k t 0 is the number of observed failures in the field or during the testPlease note that unlike the simplified Equation 13, Equation 15 can process the data with zero recorded failures. In those cases, Equation 15
38、 will still produce a non-zero value of O while k = 0 and provide any required confidence level if statistical calculations are required. Therefore, Equation 15 will be used for all FIT-rate calculations in this document.SAE INTERNATIONAL J3083 MAR2017 Page 12 of 225.3 Sources for Failure DataVariou
39、s sources can be tapped to obtain all the data required for the calculations based on Equation 15 including the following: 1. Warranty returns2. FRACAS3. Remanufacturing and/or repair data4. Dealership reports5. Other repair data (e.g., service depot, parts order records, etc.)Calculations of failur
40、e rates, O can be done based on time in service or usage (miles driven, ignition on/off cycles, number of door slams, etc.). The choice of the type of calculations should be made based on the format of the data (time versus usage). If both are available, a conversion coefficient can be calculated by
41、 linking failures per unit of usage (e.g., failures per million miles) with FIT rates.6. OPERATING TIME IN THE FIELD AND USAGE DATAIn order to successfully run the analysis, it is important to understand the assumptions and numbers used in the analysis.The time in the field is generally assumed to b
42、e the number of operating hours per day (see Table 1). However, when the failure modes are not affected by whether or not the unit is powered, then the “hours of operation per day” (second column of Table 1) becomes 24.Table 1 - Percentile vehicle usage data based on Delphi Electronics but, in a com
43、plex system, with a large number of components, these calculations can become quite cumbersome; especially if the power dissipation is variable. The snapshot below shows the formula for determining the reliability of a low dissipation, thin film resistor (based on IEC 62380 19). 0.683911()0.1 1.4 10
44、 ( ) 10 /yiii yiii iion offThSWOSWW u(Eq. 16)where:(Si)i= the ithtemperature factor related to the ithjunction temperature of the resistor mission profileWi= the ithworking time ratio of the resistor for the ithjunction temperature of the mission profileWon= the total working time ratio of the resis
45、tor with1yon iiWW Woff= the time ratio for the resistor being in storage (or dormant) mode(n)i= the ith hinfluence factor related to the annual cycles number of thermal variations seen by the resistor, with the amplitude Ti.Ti= the iththermal amplitude variation of the mission profileThe variables i
46、n this formula are dependent on the number of thermal cycles, whether or not the part is operating during each one of those cycles, the load (ratio of the operating power to the rated power) for each one of those cycles, and the actual temperature when the load is present. An electronic assembly can
47、 have hundreds of resistors, each with a different load; thus, a reliability prediction would require a formula like the one above for each resistor. As expected, the formulas become substantially more complex for more complicated components and still, the accuracy of the forecasted numbers remains
48、questionable due to variation and uncertainty. Obtaining failure rates for the specific component categories and groups without solving complicated equations similar to Equation 16 for every single electronic component on the circuit board can provide benefits from both accuracy and engineering resources standpoint. For example, the Delphi WARF database described in the case study in Section 10 of this document has 64 component categories and shows a good correlation within one order of magnitude with the field data.By making some reasonable assumptions, we can subst
