Electronics Guide

Articles in This Category

The Core Reliability Functions

Four functions describe the same underlying random variable, the time to failure T, from four different angles. Any one of them determines the other three.

Reliability function, R(t)

The probability that the item survives beyond time t, written R(t) = P(T > t). It starts at one and decreases monotonically toward zero.

Cumulative failure distribution, F(t)

The probability of failure by time t, F(t) = 1 − R(t). This is the quantity that matters for warranty exposure and field return forecasting.

Probability density function, f(t)

The derivative of F(t). It describes how failures are distributed in time across the original population.

Hazard rate, h(t)

The instantaneous failure rate among survivors, h(t) = f(t) / R(t). It answers the operationally useful question: given that a unit has worked until now, how likely is it to fail in the next interval?

The hazard rate is the most diagnostic of the four because it reveals whether a population is improving, stable, or wearing out. It also reconstructs the reliability function directly through the cumulative hazard H(t) = ∫₀^t h(u) du, since R(t) = e^−H(t). Confusing the hazard rate with the density is a common error: f(t) is normalized to the population that existed at t = 0, whereas h(t) is normalized to the units still working at time t.

The Constant-Hazard Special Case

When the hazard rate is constant at h(t) = λ, the time to failure is exponentially distributed, the reliability function reduces to R(t) = e^−λt, and the mean time to failure is simply the reciprocal of the failure rate, MTTF = 1 / λ. The exponential distribution is memoryless: a used unit that has not failed is statistically indistinguishable from a new one, so preventive replacement gains nothing. That property makes the mathematics unusually tractable, which is precisely why the assumption is applied far more often than the data justify. It is defensible during the useful-life period of a mature design and indefensible for any dominant wear-out mechanism, as the bathtub curve discussion makes concrete.

Mean Life and Percentile Life

For a non-negative time to failure, the mean life is the area under the reliability curve, MTTF = ∫₀^∞ R(t) dt. The mean is a poor summary of a skewed life distribution, so reliability requirements are more often stated as a percentile life: the B10 life (also written L10) is the time by which ten percent of the population has failed, and the median or B50 life is the time by which half has failed. Percentile lives translate directly into warranty and spares decisions in a way that a mean does not.

The Weibull distribution illustrates the point. With R(t) = exp[−(t/η)^β], the shape parameter β classifies the failure pattern: β < 1 indicates a decreasing hazard rate and infant mortality, β = 1 recovers the exponential case, and β > 1 indicates a rising hazard and wear-out. The scale parameter η, called the characteristic life, is the time at which 63.2 percent of the population has failed, whatever the value of β. The mean is ηΓ(1 + 1/β), which coincides with η only near β = 1.

Metrics for Repairable Systems

Reliability metrics divide along a line that is easy to state and frequently ignored: whether the item is repaired or discarded when it fails. MTTF applies to non-repairable items, from a single capacitor to a sealed module that is swapped out rather than opened. MTBF applies to repairable equipment and is estimated as total operating time divided by the number of failures over that time. Under a constant failure rate with negligible repair time, the two coincide numerically, which is why they are so often used interchangeably and so often misapplied.

Repair introduces a second family of measures. MTTR is the mean active repair time, covering diagnosis, replacement, and verification. Mean down time (MDT) is larger, adding detection delay, logistics delay, and administrative delay, and it is the quantity an operator actually experiences. Maintainability, treated as a design property, is what makes MTTR small: accessible connectors, modular line-replaceable units, and built-in test that isolates a fault to one replaceable item.

Combining the two families gives availability, the fraction of time a system is in a functioning state. Inherent availability, A_i = MTBF / (MTBF + MTTR), reflects only the design and excludes logistics; operational availability substitutes mean down time and mean time between maintenance actions, so it reflects the whole support system. The distinction is not academic, because a design with excellent inherent availability can deliver poor operational availability if spares are two days away.

Availability is conventionally quoted in "nines," which compress usefully into annual downtime figures. Against an 8,760-hour year, 99.9 percent availability permits about 8.8 hours of downtime, 99.99 percent about 53 minutes, and 99.999 percent about 5.3 minutes. Each additional nine costs roughly an order of magnitude more in redundancy, monitoring, and process discipline, which is why availability targets deserve the same scrutiny as any other expensive requirement. System reliability calculations and redundancy and fault tolerance describe the structures that buy those nines.

Failure Rate Units and Conventions

Component failure rates are small numbers, so the field uses scaled units. The FIT, or failure in time, is one failure per billion (10⁹) component-hours. A part rated at 100 FIT has λ = 10⁻⁷ per hour, corresponding to an MTTF of 10⁷ hours, or roughly 1,100 years, which immediately shows why such figures describe population behavior rather than the life of any individual unit. FIT rates are additive for parts in series, so a board of 500 components averaging 20 FIT each has a board rate near 10,000 FIT, an MTBF near 100,000 hours, and an annualized failure rate near 8.8 percent.

That last conversion is worth internalizing. The annualized failure rate is approximately λ multiplied by 8,760 hours when λt is small, so a component rate that sounds negligible becomes a visible warranty cost once it is multiplied by part count and fleet size. Other conventions appear in specific industries, including parts per million per year for automotive components and defects per million opportunities in manufacturing quality work.

Published failure-rate data come from several traditions, and they are not interchangeable. Handbook methods such as MIL-HDBK-217, last officially revised as Notice 2 to Revision F in 1995, sum tabulated base failure rates adjusted by quality, environment, and stress factors. Telcordia SR-332, now in Issue 4 of 2016, serves telecommunications equipment and permits laboratory and field data to be merged with handbook estimates. IEC 61709:2017, whose third edition merged the previous edition with the withdrawn IEC TR 62380, provides reference conditions and stress models for converting rates between operating environments. All are widely criticized for resting on aging component data and for implying a precision they do not possess. The prevailing view, reflected in the physics-of-failure movement and in frameworks for evaluating prediction methods such as IEEE 1413, is that handbook numbers are useful for comparing design alternatives and sizing spares, and unreliable as forecasts of absolute field failure rates. The Reliability Prediction Methods and International Reliability Standards categories treat these methods and their limits in detail.

Common Misinterpretations

Most reliability disputes trace to a handful of recurring misreadings rather than to arithmetic errors.

• Treating MTBF as a service life. An MTBF of 100,000 hours does not mean the equipment lasts eleven years. Under a constant hazard rate, the probability of surviving to the MTBF is e⁻¹, about 36.8 percent, so most units fail before reaching it. MTBF describes a failure rate during the useful-life period, not a lifetime.
• Extending the constant-rate assumption past its domain. Exponential formulas applied to a population dominated by a wear-out mechanism, such as electrolyte loss in aluminum electrolytic capacitors or thermomechanical fatigue in solder joints, understate late-life failures badly. A Weibull fit that yields β well above one is a direct warning that the exponential model does not apply.
• Quoting point estimates without confidence bounds. An MTBF computed from three failures carries enormous uncertainty. Reliability demonstration tests are therefore specified with a confidence level, and a test that ends with zero failures supports a lower confidence bound on MTBF of only T / (−ln(1 − C)) for total test time T, which is about T / 2.3 at 90 percent confidence.
• Confusing prediction with measurement. A handbook prediction is a modeling output; a field failure rate is an observation. Treating the first as evidence about the second, without closing the loop through Warranty and Service Analysis, is how programs discover their reliability problems from customers.
• Ignoring the stated conditions. Failure rates rise steeply with temperature and electrical stress. A figure quoted at 25 degrees Celsius and half-rated voltage says little about the same part at 85 degrees Celsius and full rating, which is the entire rationale for derating and for accelerated testing.
• Assuming redundancy multiplies independently. Parallel-path arithmetic assumes independent failures. Shared power supplies, shared firmware, shared calibration errors, and shared environments create common-cause failures that dominate the result once the independent terms become small.

Why These Fundamentals Matter

These metrics and methods supply the quantitative foundation for the rest of reliability practice. They let engineers set meaningful targets, size test programs, analyze field data, and trade reliability against cost and schedule with numbers rather than assertions. Equally important, a shared vocabulary prevents the misunderstandings that arise when MTBF, service life, warranty period, and design life are used as though they were the same quantity.

In program terms, reliability metrics do four jobs. They express requirements in contractual language, typically as a required MTBF, a maximum FIT rate, or a percentile life. They provide objective criteria for choosing among design alternatives and for making qualification decisions. They support warranty reserves, spares provisioning, and service planning, which is where reliability economics connects the engineering to the balance sheet. And they enable improvement by comparing what was predicted against what the field actually returns, with physics-of-failure analysis explaining the gap.

The mathematics here is demanding in places, but its purpose is practical: to make the assumptions behind every reliability number explicit, so that no formula is applied outside the conditions that justify it. An engineer who can state what a metric assumes, what it excludes, and how much uncertainty surrounds it is equipped to use the more advanced methods in design for reliability and reliability testing and qualification, and to exercise sound judgment across the product life cycle.