# Design, failure and safety: how engineering knowledge grows: Risk, redundancy and reliability numbers

Source page: https://www.materialinsights.org/learn/engineering/design-failure-safety/p5/

[All courses](https://www.materialinsights.org/learn/) / [Understanding Engineering](https://www.materialinsights.org/learn/engineering/) / [Module 12: Design, failure and safety: how engineering knowledge grows](https://www.materialinsights.org/learn/engineering/design-failure-safety/) Part 5 of 5

## Risk, redundancy and reliability numbers

Safety is finally a matter of numbers: how likely, how bad, and how many independent defences stand between a fault and a disaster. This part sets out how risk is defined, why spare parts protect less than they appear to, and how failure rates are quoted and misread.

### 5.1 Risk is the chance of a harm times its size, and each field states its target in its own units

Risk is the chance of a harm multiplied by its size. A water pump that fails once in ten years with no consequence beyond a day without water is a small risk. A bridge that fails once in a million years but kills a hundred people may count as a larger risk in the public mind, even if the product is small, because society fears large single events more than the same toll spread thin. Engineers therefore state targets for the chance of failure, matched to the consequence.

Different fields write the target in different units. The UK Health and Safety Executive's framework treats an individual risk of death above 1 in 1,000 per year for workers, or 1 in 10,000 per year for the public, as intolerable, and risk below about 1 in a million per year as broadly acceptable. In between, risk must be cut as low as reasonably practicable (ALARP), which weighs the cost of each further reduction against its benefit. Civil aviation guidance asks that a failure condition able to bring down an aircraft be extremely improbable, which means about one chance in a billion per flight hour or less. Process plants rate trip systems by safety integrity level, which states how likely the system is to fail when called on. Structural codes state a reliability index.

The table sets these side by side with a caution: they measure different things and cannot be compared line by line. What they share is the order of work. A target is set first, from the consequence, and the design is then shown to meet it by a mix of margin, redundancy, testing and inspection.

Data Targets set in different units

| Field and source | Target | What the number measures |
|---|---|---|
| Workers at a hazardous site, UK HSE framework | below 1 in 1,000 per year | individual chance of death in a year; above this is intolerable |
| Public near the hazard, UK HSE framework | below 1 in 10,000 per year | the same measure, stricter |
| Broadly acceptable level, UK HSE framework | about 1 in a million per year | below this, little further effort is needed unless it is cheap |
| Aircraft failure condition leading to catastrophe, EASA CS-25 guidance | about 1 in a billion per flight hour or less | chance per condition per hour of flight, called extremely improbable |
| Safety function in a process plant, IEC 61508 (SIL 1 to SIL 4), low-demand mode | from 1 in 10 to 1 in 100 (SIL 1) down to 1 in 10,000 to 1 in 100,000 (SIL 4) | average chance the function fails when called on |
| Structural member, EN 1990 reliability index 3.8 for 50 years | about 7 in 100,000 in 50 years (calculated), or 1.4 in a million per year (calculated) | chance of reaching the ultimate limit in the structure's life |

Conversion for the last row: the normal curve gives 7.2 x 10^-5 beyond 3.8 standard deviations; divided by 50 years this is 1.4 x 10^-6 per year. Check: 7.2 x 10^-5 / 50 = 1.4 x 10^-6.
 Source: Health and Safety Executive, Reducing risks, protecting people (R2P2) (HSE Books, 2001), tolerability of risk framework: 1 in 1,000 a year for workers and 1 in 10,000 for the public as the limits of tolerability, 1 in a million as broadly acceptable (paras 130 to 136 as commonly cited); https://www.hse.gov.uk/enforce/expert/r2p2.htm; TUV SUD, 'IEC 61508 Functional Safety', SIL table; https://www.tuvsud.com/en-us/services/functional-safety/iec-61508; Joint Research Centre, European Commission, Eurocodes presentation on EN 1990, reliability index 3.8 for 50 years; https://eurocodes.jrc.ec.europa.eu/sites/default/files/2022-06/Moscow08_9_Gulvanessian.pdf; SAS, 'EASA failure condition classifications and probability terms' (2024), extremely improbable about 10^-9 per flight hour; https://sassofia.com/blog/easa-failure-condition-classifications-and-probability-terms/  In the data

#### What one in a billion per hour means

Suppose a fleet of 500 aircraft each flies 3,000 hours a year, so 1.5 million hours in all. A failure condition held to 1 in a billion per hour then occurs, on average, 1.5 million x 10^-9 = 0.0015 times a year across the whole fleet (calculated), or once in about 670 years (calculated). Aircraft have many such conditions, so the true figure is higher.

- [Health and Safety Executive, Reducing risks, protecting people (R2P2) (HSE Books, 2001)](https://www.hse.gov.uk/enforce/expert/r2p2.htm) · web
- Kosky et al., Exploring Engineering (Academic Press, 2013) · part 3, ch. 21, 'Design Step 4. Detailed Design' (sources of risk)

### 5.2 Redundancy protects only against independent failures, and common causes usually dominate

Redundancy means more than one way to do a job, so that one fault is not a disaster. The module Control and systems showed the arithmetic: two independent pumps of 95 per cent availability give 99.75 per cent together, and a shared power feeder spoils it. This section takes that warning further, because common causes are where redundancy most often fails.

A common-cause failure is a single fault that disables several parts at once: a shared power supply, the same design error, the same maintenance slip, a flood, a fire in one room, a defective batch of parts. Reliability engineers model it crudely but usefully with a beta factor: a fraction beta of each part's failures is assumed to hit every copy together. Estimate 11 shows the effect. Two safety valves that each fail on demand once in a thousand times should fail together once in a million if independent. With only 5 per cent common cause, the joint failure rate is about fifty times worse. Beyond modest redundancy the common cause dominates, and a third or fourth copy adds almost nothing.

The remedies are separation (copies in different rooms and on different supplies), diversity (a different principle for the second path, such as a spring-loaded valve backing up an electrical trip) and, above all, independent checking of the design. Voting adds another trade-off. A two-out-of-three arrangement acts when any two of three sensors agree, so it cuts spurious trips as well as missed demands. In structures, the counterpart of redundancy is an alternative load path, so that losing one member re-routes the load. The Hyatt walkways had many hangers but a shared weak detail, so when one connection let go its load passed to neighbours no stronger. Redundancy that shares a flaw is only repetition.

Data Worked estimate 11: two safety channels with and without a common cause

| Arrangement | Chance of failing to act on demand | Spurious trips a year | Working |
|---|---|---|---|
| One channel | 0.001 | 0.01 | p = 0.001 per demand, q = 0.01 per year |
| Two channels, either can act, independent | 0.000001 (calculated) | 0.02 (calculated) | p x p = 10^-6; either channel can trip, so 2q |
| Same, with 5 per cent common cause | 0.000051 (calculated) | 0.02 (calculated) | (0.95 x 0.001)^2 + 0.05 x 0.001 = 9 x 10^-7 + 5 x 10^-5 |
| Two out of three, with 5 per cent common cause | 0.000053 (calculated) | 0.0003 (calculated) | 3 x (0.95 x 0.001)^2 + 0.05 x 0.001 = 2.7 x 10^-6 + 5 x 10^-5; spurious 3 x q^2 |

Illustrative values. Check: 5.1 x 10^-5 / 10^-6 = 51 (calculated), so the common cause makes the pair about fifty times worse than the independent estimate. Voting two out of three keeps the failure chance about the same as one out of two. The spurious-trip figure for two out of three is crude: it treats a false signal as lasting the whole year, whereas a real one lasts until it is repaired, which makes the true figure far smaller still. The point is only that voting cuts false trips sharply.
 Source: TUV SUD, 'IEC 61508 Functional Safety', functional safety method; the beta-factor model is the standard simple model of common-cause failure; arithmetic by the author

- Petroski, To Engineer Is Human (Macmillan, 1985) · ch. 8, 'Accidents Waiting to Happen' (alternate load paths)
- [TUV SUD, 'IEC 61508 Functional Safety'](https://www.tuvsud.com/en-us/services/functional-safety/iec-61508) · web

### 5.3 Reliability numbers: failure rate, MTBF and the bathtub curve

Reliability is the chance that something works for a stated time under stated conditions. For parts with a steady chance of failing at any moment, the failure rate, written lambda, is the average number of failures per hour. The mean time between failures, MTBF, is 1 divided by lambda, and the chance of surviving a time t is e raised to minus lambda times t. Electronics makers quote lambda in FIT, failures in time, where 1 FIT is one failure per billion device-hours.

Two traps catch beginners. The first is to read MTBF as a life. A part with an MTBF of 100,000 hours does not last 11 years. The formula says that at a time equal to the MTBF, 37 per cent of parts still work and 63 per cent have failed. MTBF describes the steady middle of a product's life and says nothing about wear-out. The second trap is to forget the bathtub curve. The failure rate is high at the start, when weak units, poor solder joints and installation errors show up. It then runs low and flat for the useful life, and rises again as parts wear out. Makers run burn-in tests to remove early failures, and users replace parts before wear-out begins. Constant-rate arithmetic applies only to the flat floor.

Series arithmetic is unforgiving. Estimate 12 takes a controller board of 200 parts at an average of 50 FIT. Each part is excellent, failing once in 20 million hours. Yet the board has a failure rate of 10,000 FIT, an MTBF near 11 years, and 8 per cent of boards fail within a year of continuous running. This is why complex products need either very reliable parts or redundancy, and why the cheapest part on a board often decides when a product stops.

Data Worked estimate 12: from FIT to failures in the field

| Quantity | Working | Result |
|---|---|---|
| One part at 100 FIT | lambda = 100 / 10^9 per hour | 10^-7 per hour (calculated) |
| Its MTBF | 1 / 10^-7 | 10 million hours, about 1,140 years (calculated) |
| 10,000 such parts running all year (8,760 hours) | 10^-7 x 8,760 x 10,000 | about 9 failures a year (calculated) |
| Board of 200 parts at 50 FIT each, all needed | 200 x 50 FIT | 10,000 FIT = 10^-5 per hour (calculated) |
| Board MTBF | 1 / 10^-5 | 100,000 hours, about 11.4 years (calculated) |
| Survival over one year of continuous running | e^(-10^-5 x 8,760) = e^(-0.0876) | 0.916, so 8.4 per cent fail (calculated) |
| Survival at a time equal to the MTBF | e^(-1) | 0.37 (calculated) |

Check: 1,000,000 hours is 114 years, so 10 million hours is 1,140 years. Failure rates are illustrative round figures. The constant-rate model holds only on the flat part of the bathtub curve.
 Source: NIST/SEMATECH, e-Handbook of Statistical Methods, 'The bathtub curve', three regions of failure rate; https://itl.nist.gov/div898/handbook/apr/section1/apr124.htm; arithmetic by the author  Common misreading

#### A rated life is not an MTBF

The life printed on an LED lamp pack is usually the time at which light output has fallen to a set fraction of the original, a wear-out measure. It says little about when the driver electronics may fail, which can come earlier and follows the failure-rate arithmetic above. Before comparing two reliability claims, find out what each number measures, over what conditions, and with what share of units allowed to fail.

- [NIST/SEMATECH, e-Handbook of Statistical Methods, 'The bathtub curve'](https://itl.nist.gov/div898/handbook/apr/section1/apr124.htm) · web
- Kosky et al., Exploring Engineering (Academic Press, 2013) · part 3, ch. 24, 'Design Step 7. Performance Evaluation'

### See it in the table
[H₂O Water](https://www.materialinsights.org/commodities/water/)

### Check your understanding

Two safety channels each fail on demand 1 in 1,000 times, with 5 per cent common cause. About how often do both fail together?
 The independent part is about 10^-6 but the common cause adds 0.05 x 10^-3 = 5 x 10^-5, which dominates. The total is about 5 x 10^-5, or 1 in 20,000.

A part has an MTBF of 100,000 hours and a constant failure rate. What share are still working at 100,000 hours?
 Survival is e^(-lambda t) = e^(-1) = 0.37 when t equals the MTBF. MTBF is an average rate, not a guaranteed life.

What is the main lesson of the common-cause model for a buyer who sees 'redundant' systems in a specification?
 The shared elements set a floor on the failure probability, so the question is what the copies have in common, not how many there are.

[← Part 4: The design cycle, testing and how knowledge is stored](https://www.materialinsights.org/learn/engineering/design-failure-safety/p4/)[Mark read · Wrap-up →](https://www.materialinsights.org/learn/engineering/design-failure-safety/review/)

## Other sections

- [Module overview](https://www.materialinsights.org/learn/engineering/design-failure-safety/index.md)
- [Why failure teaches more than success](https://www.materialinsights.org/learn/engineering/design-failure-safety/p1/index.md)
- [Four failures and what each one taught](https://www.materialinsights.org/learn/engineering/design-failure-safety/p2/index.md)
- [What a factor of safety really buys](https://www.materialinsights.org/learn/engineering/design-failure-safety/p3/index.md)
- [The design cycle, testing and how knowledge is stored](https://www.materialinsights.org/learn/engineering/design-failure-safety/p4/index.md)
- [Key terms, quiz and sources](https://www.materialinsights.org/learn/engineering/design-failure-safety/review/index.md)
