When field failures appear with unsettling regularity across a deployed fleet, conventional root-cause tools hit a wall. teardown and 8D discipline solve single-event manufacturing defects, but they cannot model a population that is wearing out. If 23 motors fail across 200 installed units over six months, the question is not which specific defect killed each one. The question is what failure mechanism is driving the entire curve, and how many will fail next month.
Weibull analysis answers that question. It is a statistical method for modelling time-to-failure data, named after Waloddi Weibull, who published the distribution in 1951. Unlike the normal distribution, which forces data into a symmetric bell curve, the Weibull distribution adapts its shape to the data. A single parameter, the shape parameter beta, tells you whether you are dealing with infant mortality, random failures, or wear-out.
In my own work supporting automotive and industrial customers, I have used Weibull to move quality management from reactive investigation to prediction. The analysis does not replace engineering judgment or physical root-cause work. It tells you the scope, the timeline, and the probability of what is coming — so you can act before the next failure arrives.
The Shape Parameter Tells You the Failure Physics
The power of Weibull lies in its diagnostic clarity. The shape parameter, beta, divides all failure behaviour into three categories. When beta is less than 1, failures decrease over time — these are infant mortality failures caused by manufacturing defects, poor solder joints, or assembly errors. The component either dies early or survives long-term.
When beta equals 1, the failure rate is constant. This is the exponential distribution, typical of electronic components operating under stable conditions. Failures appear random — voltage spikes, software faults, external events — with no link to operating age. When beta is greater than 1, the failure rate increases with time. This is wear-out, and it is the most critical pattern for mechanical systems: bearings, seals, gears, contactors, and insulation.
That diagnostic output is what makes Weibull different from descriptive statistics. You are not calculating a mean time between failures and hoping it is useful. You are fitting a model that reflects the physical degradation mechanism inside the component. If beta comes out at 2.7 for a mechanical assembly, you have quantitative confirmation of accelerated wear-out, and you can act on it.
The second key parameter is eta, the scale parameter, also called characteristic life. It represents the age by which 63.2 percent of the population will have failed. It is not the mean and not the median — it is a distribution-specific value that anchors the timeline. Together, beta and eta define the entire reliability curve.
Building the Analysis: Data, Ranks, and the Plot
Every Weibull analysis stands or falls on data quality. You need time-to-failure for every failed unit, measured in the relevant operating metric — hours, cycles, kilometres, not calendar time. More importantly, you need censored data: the units still in service. If you delivered 200 motors and 23 failed, the 177 survivors carry critical information about the failure distribution. Ignoring them is the most common analytical error, and it makes results catastrophically pessimistic.
Once data is assembled, failures are sorted shortest to longest and assigned median ranks using Bernard's approximation: median rank equals (i minus 0.3) divided by (n plus 0.4), where i is the failure sequence number and n is the total observation count. This formula estimates population failure probability from a limited sample — it is not a simple percentage calculation.
The data is then plotted on Weibull-scaled axes, where the X-axis represents time-to-failure and the Y-axis represents cumulative failure probability. If the points form a roughly straight line, the Weibull distribution fits. The slope of that line is beta. The point where the line crosses 63.2 percent probability gives eta. A curved plot signals mixed failure mechanisms — a warning that you must separate data by failure mode before proceeding.
Weibull Analysis Workflow
- 01Collect failure dataRecord time-to-failure for failed units plus operating hours on all surviving units
- 02Assign median ranksApply Bernard's approximation to estimate population failure probability at each point
- 03Plot and fitGraph on Weibull axes; verify linearity to confirm single failure mechanism
- 04Extract parametersRead beta for failure type and eta for characteristic life at 63.2 percent
- 05Calculate B10 and predictDerive B10 life and project near-term failure probability for the installed fleet
B10 Life: The Metric That Drives Automotive Decisions
In automotive supply chains under IATF 16949, B10 life — the age by which 10 percent of the population will have failed — is the dominant reliability metric. If your B10 falls short of the customer's specification, you will not pass the PPAP submission for safety-relevant components. Brake systems, steering modules, sensors, control units — all carry B10 targets defined in the technical agreement.
Aerospace pushes this further. Under AS9100 and EASA requirements, critical components use B0.01 — the time by which 0.01 percent of units fail. That is one in ten thousand. The calculation method is identical, but the stringency of the target transforms the engineering response. Weibull analysis combined with maintenance reliability data determines the preventive replacement interval for flight-critical hardware.
The third parameter, gamma, represents a minimum life — the point before which failure is physically impossible. The three-parameter Weibull model applies when a genuine warranty period or physical barrier exists. Use it cautiously: forcing gamma into a dataset where it does not belong will distort beta and eta. Test the two-parameter model first and compare goodness-of-fit.
Key Weibull Reliability Metrics
From Diagnosis to Prediction: A Worked Case
I was brought into a case involving electric motors failing on a Swiss production line with unsettling regularity. Three failures in two weeks became five in a month. No single manufacturing parameter explained why specific units died. When I plotted the failure data — 23 failed motors with precise operating hours, plus 180 survivors still in service — the Weibull output was clear and actionable.
Beta came out at 2.7, confirming wear-out failure mode. Eta was 14,300 operating hours. B10 sat at 5,200 hours. The customer ran these motors continuously, which meant B10 was reached in approximately 217 days — just over seven months of deployment. That timeline matched the failure reports precisely. The model was not theoretical; it was describing what was happening on the floor.
The prediction was what changed the customer's posture. With 180 motors in operation, the model calculated a probability exceeding 75 percent that 8 to 12 additional units would fail within the next 30 days. That number moved the conversation from quality investigation to executive risk management. The decision was made immediately: preventive replacement of all motors exceeding 4,500 operating hours.
Weibull does not just tell you what failed. It tells you what fails next, when, and with what probability — and that changes the decision.
Simultaneously, the supplier was engaged to redesign the bearing system. The revised motor achieved B10 of 11,800 hours — more than double the original. The customer never saw the Weibull plot, never discussed beta or eta. They saw motors that performed as specified, and a quality organisation that flagged the problem before it escalated.
Integrating Weibull with Core Quality Tools
Weibull analysis does not operate in isolation. In a mature quality system under IATF 16949 or AS9100, it feeds directly into PFMEA. Weibull-derived failure probabilities and B10 values provide quantitative inputs for severity and occurrence ratings that would otherwise rely on subjective scoring. A wear-out failure mode with a confirmed beta of 2.7 and a B10 inside the warranty window carries an unambiguous risk priority.
For system-level reliability, Weibull parameters feed into Reliability Block Diagrams. Individual component eta and beta values allow calculation of overall system reliability for series and parallel arrangements. This is how you determine whether a redundant sensor architecture genuinely meets the availability target, or whether the weakest link pulls the system below its required MTBF.
Preventive maintenance intervals are the most direct application. Weibull identifies the inflection point where the failure rate begins climbing sharply. The optimal replacement window sits just before that acceleration — not so early that usable life is discarded, not so late that unscheduled downtime dominates. Warranty cost analysis follows the same logic: project field failures against claim costs and set the warranty boundary where exposure is controlled.
Pitfalls and Practical Starting Points
The most dangerous trap in Weibull analysis is mixed failure mechanisms. If your dataset contains bearing failures and electrical overstress events, the plot will curve and beta becomes meaningless. You must classify failures by physical mechanism before fitting the model. Run separate analyses for each mode, then assess the combined population risk. This is non-negotiable.
Sample size limits confidence. Weibull is mathematically robust with as few as five or six failures, but the confidence intervals around beta and B10 will be wide. Reporting a single B10 value from a small dataset without its confidence bounds is misleading. Always state the range, and make sure management understands the difference between a point estimate and a prediction interval.
Extrapolation beyond the data is a professional risk. If your longest observed failure occurred at 10,000 hours, any prediction about 20,000-hour behaviour carries significant uncertainty. The model can extrapolate — the mathematics allow it — but the engineering responsibility does not. Report confidence intervals at every extrapolated point and qualify the prediction explicitly.
Start with one component family where you have at least 10 documented failures and a known population of surviving units. Use established software — Minitab, ReliaSoft Weibull++, or JMP — and verify the plot manually the first time. Cross-check beta against the known failure physics of the component. If the statistics contradict the engineering, one of them is wrong, and resolving that contradiction is where the real insight lives.