August_EDFA_Digital

edfas.org 1 1 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 22 NO. 3 in field shape knowledge (ppm at each time t ). Anyway, estimationmethod choice is highly dependent on context and number of field returns. Indeed, two typical contexts to estimate a latent failure rate may be faced. For a new product, at the design and qualification step and before manufacturing, reliability tests, such as high temperature operating life (HTOL) tests, are run to accelerate product life and record failures at test end (test duration is calibrated to correspond to real life duration). In a typical casewhere nomonitoring is performed during the reliability test, it is not possible to study failure rate evolution during product life, so failure rate is assumed constant: in field modeling, it would be the exponential model case. A confidence interval is built on this constant failure rate (ppm/year between low and high values at Alpha risk). As a second context to estimate a latent failure rate, when manufacturing is started and parts are running in cars, a real field modeling can be implemented, the read points that missed in HTOL tests to model failure rate in reliability laboratory, fitting there with the car failures reported by mechanics. [1] So, from field returns seen as read points, it becomes possible to model field shape in time. Obviously, if failure events are not reported in time or if only a few failures are reported, field modeling is not really implementable and a risk assessment will be built only on a failure rate estimation in time, without modeling. Anyway, in all the field failure cases, when field modeling is possible or not, confidence intervals will be built, on the constant failure ratewithout modeling, or on model parameters and failure rate at each time in case of field modeling. A field rate is always fitting with a com- parison between a number of failures ob- served at time t andquantity of likely-to-fail parts in the field reported as not having failing yet at this same instant. Confidence interval width is typically a function of risk level accepted and inversely proportional to sample size, that is corresponding here with quantity of surviving parts: in the field modeling case, confidence interval will be also enlarged if only a few field failures are reported. Confidence interval width is directly prediction accurate, so that it becomes important to act to reduce this interval, using larger samples when it is possible. In qualification context, before manufacturing and life in field, failure rate estimation performed from HTOL test results takes into account other test results performed from similar products to the directly tested one, on a similar design or technology basis. When a failure is observed in reliability, always in this same qualification step, it will be corrected, and failure rate estimation will be performed on a population that could have also this failure but for which no same failure was reported. If a correction of this defect is not possible, for examplewhen it deals with a defect density, some different estimation methods can be proposed. [2] In field context, confidence interval is usually smaller since all the manufactured parts in field can be used in failure rate estimation. Nevertheless, including some dif- ferent failures or different products could be studied to still reduce confidence interval. Merging different failures in a fieldmodeling can be interestingwhen it deals to have some knowledge on a customer’s application, but each specific field shape fitting with each failure will be lost in this merging (see Fig.1). A risk assessment is typically performed for only one failure, so that field returns from different failure types are definitively not mixed together. A RISK ASSESSMENT IS TYPICALLY PERFORMED FOR ONLY ONE FAILURE, SO THAT FIELD RETURNS FROM DIFFERENT FAILURE TYPES ARE DEFINITIVELY NOT MIXED TOGETHER. Fig. 1 A global field modeling performed on two different products A and B, merged together, may get each specific field shape lost.