Maintenance and Reliability

How Do You Calculate Failure Rate?

Failure rate is easy arithmetic and easy to misuse. The formula assumes a constant hazard, and the two life stages where that assumption fails are the two you most often care about.

Failure rate, usually written as lambda, is the number of failures divided by total operating time across all units observed. Take 50 units run for 1000 hours each, which is 50,000 unit-hours, with 4 failures. Lambda is 8.00e-5 failures per hour, and its reciprocal, mean time between failures, is 12,500 hours. In electronics the same figure is usually quoted as a FIT rate, failures per billion hours, which here is about 80,000.

• Failure rate in the worked example: 8.00e-5/h
• MTBF, its reciprocal: 12,500 h
• FIT (failures per billion hours): 80,000

The assumption hiding inside the arithmetic

That calculation assumes the failure rate is constant over time, which is true only during the middle of a product's life. The classic bathtub curve has three regions: early failures from manufacturing defects, where the rate starts high and falls; a long flat useful-life region, where a constant lambda genuinely holds; and wear-out, where the rate climbs. Computing a single lambda across data spanning all three produces a number that describes none of them, and it will systematically mislead about the period you actually care about.

Why MTBF gets misread as a lifetime

An MTBF of fifty thousand hours does not mean units last five and a half years. It means that within the useful-life period, across a population, failures occur at that average spacing. Individual units fail at all sorts of times, and a high MTBF is entirely compatible with a product that begins wearing out well before the MTBF figure. Confusing the two leads to warranty terms set from a reliability number that was never a lifetime prediction.

MTBF is a rate expressed in units of time, which is exactly why it keeps getting mistaken for a lifespan.

How much data before the number means anything

Failure counts are small integers, and small integers carry wide confidence intervals. With four failures, the true rate could plausibly sit at roughly half or nearly double the point estimate. This is why reliability work reports confidence bounds rather than a bare figure, and why comparing two suppliers on a handful of failures each is usually comparing noise. If a decision rests on the difference between two failure rates, check whether the intervals overlap before treating the gap as real.

Censored data is the norm, not the exception

In most real datasets, some units are still running when the observation ends. Those are censored observations, and discarding them biases the estimate badly because they are disproportionately the reliable units. The simple formula above handles this correctly only if you count their accumulated hours in the denominator while not counting them as failures, which is precisely what the unit-hours approach does. Counting only failed units, a common shortcut, produces a far more pessimistic answer.

Use the failure rate calculator to convert unit-hours and failure counts into lambda and MTBF. Run your failure data

Published 2026-08-08.