Field Failure Data, Weibull Analysis and Decisions
Field failures are the only data that describes what the product actually does, and they arrive in small numbers with incomplete information. The analysis has to make a decision from limited evidence.
Why Field Data Is Different
The units are in an environment that was not controlled, at an age that varies, and the failure may be reported by someone who is not an engineer. The data is a report rather than a measurement.
Despite that, it is the only information about the failure rate the customer experiences. Our failure analysis notes describe how a returned unit is handled.
Getting Usable Data
The minimum for an analysis is the number of units in the field, the number that failed, and the time each was in service. Without the population, a failure count cannot be converted into a rate.
The report should record the conditions, the symptoms and the disposition, so that failures from a different cause are not merged into one set. Our yield notes describe the classification discipline that applies.
Distributions and Their Meaning
A decreasing failure rate over time indicates an early life population, which points at a manufacturing defect. A constant rate indicates random failures, which usually points at the environment.
An increasing rate indicates wear out, which points at a mechanism with a finite life such as thermal fatigue. The three are distinguished by the shape and not by the average. Our thermal cycling notes describe the test that reproduces the third.
Weibull Analysis
The Weibull distribution describes the life of a component with a shape parameter that indicates which of the three regimes applies and a scale parameter that indicates the characteristic life.
The confidence interval matters as much as the estimate, because a field sample is usually small. A point estimate without an interval overstates what the data supports.
Comparing to the Requirement
The requirement is usually expressed as a failure rate at a time, or as a life. The estimate from the field data should be compared with it, with the confidence interval stated.
Where the interval spans the requirement, the data cannot confirm or refute it, and more units or more time are needed rather than a conclusion.
Acting on the Result
Where the rate is above the requirement, the mechanism should be established from the returned units rather than from the distribution alone. The distribution says when, and the analysis says why.
The corrective action should be verified in the field, which means the rate after the change has to be measured over a comparable period. Our change control notes describe how the change is recorded.
Process Control and Verification
Running a first article through the same checks as the production panel confirms that the two agree, and that comparison is the cheapest form of process control available at prototype stage. Keeping a sample from the panel turns a dispute into a measurement, because the same coupon can be re-examined by both parties without rebuilding the batch.
Reviewing the design before the data is released is cheaper than correcting it after the panel is in the tank, because every step downstream inherits the decision made at the front end. Documenting the assumption is part of the design work, and a short note on the drawing prevents a question that would otherwise arrive a day later and cost a day of schedule.
The process window is set by the narrowest step in the flow, so an improvement anywhere else shows up as margin rather than as yield until that step is addressed. A short note on the drawing about handling, storage or packaging is often worth more than an extra decimal place on a tolerance.
Where a value sits close to a process limit, the drawing should say so, since the shop can then open the process window rather than working to a nominal figure that carries no tolerance.
Checks Before Release
The checks that matter are the ones performed on the product rather than on a sample kept for the purpose, because a coupon that travels with the panel is the only evidence about that panel. Where a requirement can be measured, it should be measured at the point of manufacture and recorded against the board or the lot it applies to.
A parameter that is set once and never re verified drifts, and the drift is usually discovered by a defect rather than by the record. The tooling, the material and the profile form one system, and a change to any of them should be assessed against the other two before it is released.
Where the process window is narrow, the measurement resolution has to be better than the window, or the data cannot distinguish a good part from a marginal one. A record that identifies the operator, the date and the settings is worth more than a record that identifies only the result.
The acceptance criteria should be written before the work starts, so that the decision is made by the specification rather than by the person inspecting. Where an operation cannot be verified afterwards, it has to be controlled during the operation, and that control has to be visible in the record.
Verification and Records
Sampling is a compromise between cost and confidence, and the sample size should follow from the failure rate that has to be detected. The first article confirms that the setup matches the intent, and it is the cheapest point at which a wrong setup can still be corrected.
Handling between operations is part of the process, and the damage it causes is often attributed to the operation that preceded it. Where two operations share a tolerance, the allocation between them should be explicit rather than left to whichever is measured first.
FAQ
How many failures are needed? A trend needs a handful and a rate estimate needs more. The number determines the confidence and it should be stated rather than assumed.
Can a field failure be analysed without the unit? The symptom can be classified and the mechanism cannot. A report is not a substitute for the unit.
What does gopcba provide for field data analysis? We provide a population based rate rather than a failure count, classification by cause and condition, distribution shape interpretation for early life, random and wear out mechanisms, Weibull estimates with confidence intervals, comparison with the requirement, and a corrective action verified by the subsequent field rate.
