Using process data to predict Product Failure Rates
October 15th, 2016
It is well established that Process Failures correlate closely to those found in the Early Life operational stage of an Electronic Product, but how do we take the failure rate or yield data from a process and convert it into an accurate failure rate prediction?
With high volume consumer electronic products, data shows 70-80% are process related, hence the higher failure rates in first 6 to 12 months of a product’s life are caused by the Early Life distribution that basically ‘sits on top’ of the Design failure distribution which is more constant and is often classed as the ‘Intrinsic’ failure rate.
Early Life Failure stress testing can of course be used to stimulate the defects and enable resolution, but of course this takes time to build up sufficient volume testing and find wide enough range of defects. This therefore makes accurate prediction difficult with much sample error and statistical variation in predictions.
It is also very difficult if not impossible to cover all possible Failure mechanisms in new product Accelerated Stress Testing which is good at detecting possible Epidemic cases, but even these are also seen at some point within process failure data.
DPPM levels of Individual Failure Types are also so low, Accelerated Testing will only address epidemic and Very Major Failure Types, hence the need for a model that focuses on using Volume Production Process Yield data. Different Production Vintages will have different Process Yield levels due to Material, Manpower, Method changes and this will lead to differing process Escape Levels and varying Early Life Field Failure levels.
There is however an approach that can greatly benefit the Reliability Engineer in trying to estimate the level of Early Life Failures that are likely to occur in the field which does not rely only on stress test failure data
This approach involves a series of steps;
1. Compare the defect types / components / location, etc found in manufacturing process with those found in the field to measure the correlation
2. If good correlation, take the overall process ROLLED yield data that represents the fraction of production that is totally defect free
3. Use different prediction models and assess the best fit model that gives best output which relates closely to actual field failure levels experienced in the field
4. Set up prediction model using rolled yield data for the sub-assembly or complete assembly and use as a key reliability indicator
Component manufacturers have used models to predict field fail levels from process yield data for many years, one of most common methods is;
Defect rate = 1 – Y(1-t ) / t
Where Y = End of line Test Yield
t = Test Coverage If Yield = 99.5% , Test Coverage = 95% , Escape Defect Rate = 263 ppm
Reliability Solutions have developed models for a range of consumer type products and sub-assemblies, one of the most successful has been to apply the equation below to populated Power Supplies;
= ((1-(R9^((1-0.838)/0.838)))100) *R9 – Rolled Yield
This provides output for individual production month
Using Process Data to Predict Product Failure Rates
Such predictions are extremely powerful for the Reliability Engineer and can be used to then make trend analysis and compare directly to similar vintage field performance.
When defect types are same, it is obvious that we should expect good correlation with field data if the chosen prediction model is suitable, correlation of failure locations is very obvious in the table below.
Also it can be clearly noted that the standard High Temperature Ongoing Reliability Test (ORT) rarely finds the same defects as each individual defect ppm level is so low and also the Defect Detection Capability of simple High Temperature testing is extremely weak.
The field failure trend chart below shows this to be the case where the Power Board failure prediction is a good fit to what is recorded as Field Return rate (FRR) each month.
In the above case we might argue the prediction model data fit constants could be adjusted to give an even better fit, this must always be reviewed to include prediction accuracy as more accurate field data becomes available.
Summary Process failure data is FREE to the Reliability Engineer and it is incredibly wasteful not to utilise it to enhance Failure rate Prediction methodology and use the varying process yield data to monitor variability in the predicted Early Life Field Return Rate.
This approach can become VERY POWERFUL, especially if managing suppliers who have limited ability for Early Life Reliability stress testing to stimulate latent defects.
Martin Shaw, Wilde Associate at Reliability Solutions