August_EDFA_Digital

edfas.org 13 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 22 NO. 3 Fig. 2). Sometimes, while risk assessment is performed far before maximum failure rate event, modeling may miss maximum failure rate existence and model providing the best fitting may wrongly not be a DS one yet, at the risk assessment date. Unfortunately, a bad risk assessment predicting a low failure rate thatmay be revealed farmuch higher later, is as serious as a failure rate predicted high at the first assessment. A risk assessment is always followed by actions linked to the predicted failure rate value (final shipment of suspectedparts if failure rate is predicted low, or car recall in the contrary case), and these actions may be finally counterproductive if predicted failure rate is fully wrong. Unfortunately, this specific risk due to competitive time and mileage actions in this case is not fully covered by typical confidence intervals andneeds tobe considered in itself differently as presented in the following. WHEN DATA FROM THE FIELD IS NOT AVAILABLE YET Previously field modeling was performed separately for each product. Now, consider the case if only one field modeling is run on all the datamerged for all the products impacted, failing and surviving parts: the purpose is to obtain a best prediction by smaller confidence intervals, as similar product families are merged when it deals to estimate a failure rate from HTOL test results. A second study goal would be to find a methodology to be imple- mented to be able to predict failure rate, from this global model, for another product impacted for which no data would be available yet. A first result obtained from the unique global field modeling is that still the same DS Frechet model provides the best fitting on data, which is indicated by a smaller Akaike’s Information Criterion (AICc) and a smaller Bayesian Information Criterion (BIC), used respectively to comparedata explanationquality andpredictionaccuracy (see Fig. 3). A second result shows that confidence interval on this global model is, as predicted, far smaller than the ones for each elementary model (width is reduced by 80%). So, confidence interval for the global model cannot cover all variability for all the different models, which makes its usage not possible to predict failure rate for another product, except adding some other weighting coefficient. Conclusion about confidence interval width would be the same one, using intervals on the events (pointwise intervals) insteadof intervals on themodels (simultaneous intervals) as typically used. The next mission lies in determining this weighting coefficient that would come to increase or decrease risk, and to implement it on the global model in order to predict failure rate for any other impacted product (see Fig. 4). More precisely, the following assumption is tested: a difference in product design (edge seal design and number of metal layers) acts on sensitivity-to-defect in termof failure quantity (parameter p), while a difference inproduct application (application tempera- ture) makes the maximum failure rate event more or less early in time. So, design of each product is checked, each customer applica- tion is studied and, finally, elementary field models are located relatively to the global field model (parameter p and max failure rate time). At 3 years, global field model is themedianmodel in termof prediction; at 15 years, it shows a higher failure rate than the one for 30% of products. Additional reliabil- ity tests are performed on some products, to check application temperature impact: these Fig. 2 Modeling far before maximum failure rate time, the maximum failure quantity may be underestimated. Fig. 3 Global field model: model comparisons and DS Frechet parameters and probability function.