Reliability Engineering: MTBF, Weibull, and Life Data
Reliability Engineering: MTBF, Weibull, and Life Data
Joshua R. Lehman
Author
Failure Analysis13 min read
Engineering reliability is the probability that a product performs its intended function under specified conditions for a specified period. This definition has three components that matter equally: the function (what the product must do), the conditions (the environment and loading it must operate in), and the time (how long it must perform). A reliability statement without all three components is incomplete.
Reliability engineering provides the quantitative tools for measuring, predicting, and designing to achieve a specified reliability target. It translates field failure data into actionable design parameters, turns warranty return rates into population-level behaviour patterns, and enables evidence-based decisions about maintenance intervals, design life, and acceptable risk. The alternative — designing by intuition and discovering reliability deficiencies in service — is substantially more expensive and slower to correct.
Reliability is expressed as a probability, R(t), the probability that a unit from a population will survive to time t without failure. A reliability of R(10,000 hours) = 0.95 means that 95 percent of units are expected to survive to 10,000 hours of operation.
The complementary quantity is the cumulative failure probability, F(t) = 1 − R(t). F(10,000 hours) = 0.05 means that 5 percent of the population is expected to have failed by 10,000 hours.
The failure rate, h(t) — also called the hazard rate — is the instantaneous probability of failure per unit time for a unit that has survived to time t. Failure rate is not constant for most real products. It changes over the product's life in a characteristic pattern described by the bathtub curve.
Reliability Is a Population Property, Not a Component Property
Reliability describes the behaviour of a population of nominally identical
units, not the specific lifetime of any individual unit. A component with
R(5,000 hours) = 0.90 does not mean every unit will operate until exactly
5,000 hours and then fail. It means that, across a large population,
approximately 90 percent of units will still be operating at 5,000 hours. Some
will fail much earlier; some will last far longer. Population-level statistics
apply to planning and design decisions; individual unit failure is not
predictable from them.
Mean time between failures (MTBF) is the expected average time between failures for a repairable system, or — for non-repairable components — the mean time to failure (MTTF). MTBF is the reciprocal of the failure rate when failure rate is constant: MTBF = 1 / λ.
MTBF is widely used in reliability specifications and procurement requirements because it is simple to understand and easy to quote. A pump with an MTBF of 50,000 hours is intuitively more reliable than one with an MTBF of 10,000 hours.
However, MTBF obscures critical information about the shape of the failure distribution. Two products can have identical MTBF values while having completely different reliability profiles. One product with MTBF = 20,000 hours may have a constant failure rate — failures are random and age-independent, common in electronic systems. Another product with the same MTBF may have an increasing failure rate — early life is very reliable but failure rate climbs steeply after 15,000 hours, producing a cluster of failures in a narrow time window. The first product is managed with random replacement or condition monitoring; the second requires scheduled replacement before the failure rate rises steeply. MTBF alone does not reveal which regime applies.
MTBF Does Not Mean Half the Units Survive to MTBF
A common misunderstanding: if a component has MTBF = 10,000 hours, then 50
percent of units have survived at 10,000 hours. This is only true if failures
follow an exponential distribution (constant failure rate). For increasing
failure rates — which characterise mechanical wear-out — the median life is
substantially less than the MTBF. A population with MTBF = 10,000 hours and an
increasing failure rate may have 30 percent of units failed by 10,000 hours.
Specifying MTBF for mechanical components without the associated Weibull
parameters produces unreliable predictions.
The bathtub curve describes how failure rate typically varies over a product's life in three distinct phases.
Infant mortality (early failure) period. Failure rate is initially high and decreasing. Failures in this phase are caused by manufacturing defects, substandard materials, assembly errors, and design weaknesses that cause some fraction of the population to fail quickly. High early failure rates indicate quality control problems in manufacturing or an inadequate design verification process. Burn-in testing — operating units under accelerated conditions before shipment — screens out infant mortality failures before products reach customers.
Useful life (random failure) period. After early failures are screened out, failure rate drops to a lower, approximately constant level. Failures in this phase are random — caused by unexpected loading events, environmental extremes, or chance occurrences rather than systematic degradation. Constant failure rate is the defining characteristic of this phase; it is well-modelled by the exponential distribution. Many electronic systems and some mechanical systems spend most of their intended service life in this region.
Wear-out period. Failure rate increases as accumulated damage — fatigue, wear, corrosion, or creep — reaches a threshold in an increasing fraction of the population. Mechanical systems with fatigue or wear failure modes typically show a pronounced wear-out region. Scheduled maintenance and replacement intervals are calibrated to keep the population out of the wear-out region during service.
The Weibull distribution is the standard model for life data in reliability engineering because it can represent all three regions of the bathtub curve depending on its parameters. It is defined by two parameters: the shape parameter β (beta) and the characteristic life parameter η (eta).
The shape parameter β controls the failure rate behaviour. When β < 1, failure rate decreases with time — infant mortality. When β = 1, failure rate is constant — the exponential distribution, representing random failures. When β > 1, failure rate increases with time — wear-out. The higher the value of β above 1, the more rapidly failure rate increases (the steeper the wear-out). Typical values for common failure modes:
β ≈ 0.5: severe infant mortality (electronic components with screening failures)
β ≈ 1.0: random failures (electronic systems, some hydraulic systems)
β ≈ 1.5–2.5: early wear-out (rolling element bearings, some seals)
β ≈ 3–4: normal wear-out distribution (mechanical components with regular maintenance)
β > 5: very tight wear-out — nearly all units fail within a narrow time window (fatigue with consistent stress history)
The characteristic life η is the time at which 63.2 percent of the population has failed. This is a mathematical property of the Weibull distribution and not directly a design target, but it scales the time axis. A larger η indicates a longer-lived population.
Beta Greater Than 3 Justifies Scheduled Replacement
When Weibull analysis of field data produces β > 3, the failure distribution is tight enough that scheduled replacement before the characteristic life makes practical sense. With β = 1 or β = 2, the spread of failure times is so wide that scheduled replacement removes many units that had substantial remaining life. With β > 3 to 4, the failure times cluster tightly enough that scheduled replacement near the characteristic life captures most of the population's useful life while preventing wear-out failures in service. This is the quantitative basis for planned maintenance intervals on components like brake pads, drive belts, and seals.
A Weibull plot graphs cumulative failure probability on a special probability scale against time on a logarithmic scale. If the data follows a Weibull distribution, the points lie on a straight line. The slope of that line is β; the intercept (at 63.2 percent failure probability) is η.
Reading a Weibull plot:
Steep slope (high β): tight failure distribution, wear-out dominated. Most units fail within a narrow time window. Scheduled replacement is effective.
Shallow slope (low β near 1): broad failure distribution, random failure dominated. Failures occur across a wide range of times. Condition monitoring is more effective than scheduled replacement.
Concave curve or two distinct slopes: mixed failure modes. The early portion of the curve represents one failure mode (possibly infant mortality or a defect-related failure); the later portion represents a different failure mode (wear-out). Two-parameter Weibull fitting to the full dataset misrepresents both modes. Analysis requires separating the populations or using a three-parameter or mixed Weibull model.
When Weibull analysis is conducted on field warranty data, the results directly indicate whether the current design is meeting its life targets and what type of corrective action is needed. A β < 1 result from field data points to manufacturing quality or early design issues; a high-β wear-out result with η below the target life indicates a design margin problem that requires redesign, material upgrade, or maintenance interval adjustment.
Weibull analysis requires life data — the ages at failure for a sample of units, plus the ages of units still operating (suspended, or censored data). Both types of data are essential. Analysing only the failures and ignoring the survivors dramatically underestimates the characteristic life and produces an overly pessimistic reliability picture.
Field data is the most accurate source for Weibull analysis because it reflects actual service conditions, actual use cycles, and actual environmental exposure. Accelerated life testing produces life data faster by applying elevated stress (temperature, load, humidity), but the stress-life relationship must be validated to translate accelerated test results to field life predictions.
Censored Data Is Not Missing Data
In life data analysis, a suspended unit — one that has not yet failed when the
analysis is conducted — is not a missing data point. It is a data point with a
known lower bound: the unit survived at least to its current age. Censored
data carries real reliability information and must be included in the
analysis. Analysis methods such as median rank regression and maximum
likelihood estimation handle censored data correctly. Discarding suspended
units from a Weibull analysis produces a systematic bias toward pessimistic
(lower) life estimates.
Minimum sample sizes for meaningful Weibull analysis depend on the expected shape parameter and the desired confidence interval. In practice, fifteen to twenty failure data points with an equal or greater number of suspensions produces reasonable parameter estimates. Fewer than ten failures produces wide confidence intervals that limit the practical utility of the results — though even small samples provide directional information about β and the failure mode type.
A hydraulic gear pump used in an industrial press was being replaced at a fixed interval of 4,000 hours based on the manufacturer's published MTBF recommendation. The maintenance team had accumulated failure records on forty-seven pumps over three years, including both failed units (twenty-two pumps) and units still in service at the time of the analysis (twenty-five pumps still running, ranging from 800 to 5,200 hours).
Weibull analysis of this dataset produced β = 2.7 and η = 6,200 hours. The R(4,000 hours) calculated from these parameters was 0.82 — meaning 18 percent of pumps were expected to fail before the 4,000-hour replacement interval. The current replacement interval was not preventing wear-out failures; it was removing pumps well before many of them would fail (at 4,000 hours, only 18 percent had failed) while still allowing nearly one in five units to fail before replacement.
The β = 2.7 value confirmed a wear-out failure mode with moderate spread. The tight-enough distribution to support scheduled replacement, but the characteristic life of 6,200 hours indicated the current 4,000-hour interval was conservative for most pumps. Examining the failure records for the twenty-two failed pumps showed that fourteen had failed between 5,000 and 7,500 hours — all in service longer than the 4,000-hour replacement interval because they had been deferred.
The deferred replacements were creating the in-service failures. The failure population was dominated by units that had slipped past the scheduled interval, not by the design life of the pump itself.
Two actions were taken. First, the maintenance scheduling system was revised to flag pumps due for replacement at 3,800 hours with a hard limit of 4,200 hours — eliminating the informal deferrals that had been allowing pumps to run past their interval. Second, a condition monitoring check (fluid analysis and vibration monitoring) was added at 3,000 hours to identify pumps showing early deterioration and allow targeted early replacement of the highest-risk units.
After twelve months, zero pump failures in service were recorded across the affected press fleet. The Weibull analysis had identified that the failure mode was schedule compliance, not an inadequate replacement interval.
Weibull parameters extracted from field data or accelerated testing become design inputs for the next product generation. A measured β value identifies the failure mode type; a measured η value calibrates the design life prediction. When the measured life falls short of the target, the gap defines the design margin improvement required.
Increasing the characteristic life η requires reducing the damage rate — using higher-strength materials, improving lubrication, reducing operating stress, or tightening manufacturing tolerances that produce early-failure variation. Increasing β (tightening the spread of failure times) requires reducing variability in the failure mechanism — tighter material property control, more consistent manufacturing, more uniform operating conditions.
Reliability targets should be specified in Weibull terms when possible, not just as MTBF. A target of "R(10,000 hours) = 0.95 with β ≥ 2.5" is a more precise and actionable reliability specification than "MTBF ≥ 100,000 hours" — it specifies both the survival probability at the target life and the expected failure mode character, and it can be verified directly against measured Weibull parameters from test data.
Life Data Turns Field Experience into Design Knowledge
Every field failure is an expensive data point. A company that collects
failure age data systematically — records the operating time at failure for
every warranty return, every field replacement, every maintenance removal —
builds a life data asset that makes every future design decision more
accurate. The first Weibull analysis on a component may be based on limited
data with wide confidence intervals; the fifth analysis, incorporating three
additional years of field data, produces tight parameter estimates that
directly improve maintenance interval setting, warranty cost prediction, and
next-generation design targets. The discipline required is consistent data
collection, not sophisticated analysis.
Reliability engineering quantifies how a design performs against its life targets using field data and statistical tools. The next post closes the loop from analysis to design: design for reliability applies these quantitative insights proactively, during the design phase, to build robustness into products before they reach the field. The key methods are stress-strength analysis, design margins, and accelerated life testing — the tools that translate a reliability target into design decisions.
Reliability is the probability that a product performs its function under specified conditions for a specified time — all three components are required for a complete reliability statement
MTBF describes the average but hides the shape of the failure distribution — two products with identical MTBF may have completely different failure rate behaviour and require different management strategies
The Weibull shape parameter β identifies the failure mode type: β < 1 indicates infant mortality, β = 1 indicates random failures, β > 1 indicates wear-out; higher β means tighter wear-out clustering
Weibull analysis requires both failure data and suspension data (units still running) — discarding suspended units produces a systematic bias toward pessimistic life estimates
Reliability targets specified as R(t) with a minimum β value are more actionable than MTBF targets because they specify both the survival probability and the expected failure character
