Predict a design's failure rate from its parts list

A reliability prediction answers the question every new design gets asked before it has run an hour: how often will it fail? With no failure data yet, the only honest answer is an estimate from published base rates, and that's what Modelling → Reliability prediction gives you: a parts count.

What a parts count is

Every part type in a handbook has a base failure rate (λb): failures per million hours for a ceramic capacitor, a film resistor, a rectangular connector, a MOSFET. Reliafy's reliability data library carries these rates from MIL-HDBK-217F, the U.S. military's prediction handbook, each one cited (browse them under Models › Library).

For microcircuits (logic, microprocessors, memories, GaAs devices), motors, synchros and resolvers, circuit boards, quartz crystals, lamps and laser diodes, the library carries the handbook's parts-count rates instead (its Appendix A): a generic failure rate (λg) for each environment, with the environment already in it. Microcircuits come in complexity bands (gates, transistors, bits or memory size), so pick the band your part falls in.

A parts count adds them up. For each line of your parts list:

failure rate = λb × πE × πQ × quantity

• λb is the handbook's base rate for that part type.
• πE is the handbook's environment factor: the same part fails more often in a helicopter than in a lab, so you pick where the equipment works (ground benign, ground fixed, ground mobile, naval, airborne, and so on) and every line uses its own factor for it.
• πQ is a quality factor for the part's grade, which you set per line. Left blank it's 1, except for the parts-count types, which take the lowest grade the handbook lists: 10 for a commercial microcircuit (class B is 1, class S 0.25), non-MIL for a crystal or a board. Each part type's grades show on the box.

For a parts-count type the line is simply λg × πQ × quantity: the handbook worked out λg from its full model at its default temperature, package and complexity, and the record says which defaults it holds.

The lines' rates add up (every part in series, each with a constant rate), and from the total you get the MTBF (1 / failure rate) and the reliability over a mission time (R(t) = exp(−λt)). A Pareto shows which lines and which part categories drive the total, and if your lines sit in sub-assemblies, the rate per assembly too.

Make one

1. Open Reliability prediction and press New prediction.
2. Give it a name, pick the environment, and set the mission time in hours. The reliability figure is the chance of no failure over that long.
3. Build the parts list. Paste a list takes one part per line with its quantity after a comma (Ceramic capacitor, 48, Flight controller; the third column is the assembly it's in); Add a line adds one at a time.
4. Give each line a handbook part type: type a word or two (capacitor ceramic, resistor film, microprocessor MOS, memory DRAM, motor, crystal) and pick the match. Set the quality factor (πQ) on a line when its grade isn't the default, or choose its grade under Grade and environment… (Environment… for a part type with no grades), which also sets an environment for that line alone. These are the part type's covariates, the same inputs as on its library page.
5. Press Calculate, read the result, then Save prediction.

The Free plan keeps 1 saved reliability prediction (the shared sample doesn't count); calculating without saving has no limit. To save another, delete the one you have or upgrade to Pro, which removes the limit.

MIL-HDBK-217F covers electronic parts. For a mechanical part (a bearing, a seal, a pump, a spring, a gear), the search also lists the Navy's NSWC-11 handbook, under its own heading:

• Pick an NSWC-11 part type, then open Conditions… on the line. Each input shows its unit, its default and the handbook's limits; the rate updates as you type, and a value outside the limits says which one and on what page. A part whose rate is per million cycles also needs how many cycles it makes an hour. A line missing a condition is left out of the total until it has it, and the result says so.
• The result's method names the lines that used each handbook. CDNSWC and the Navy did not take part in or approve Reliafy, and the result says so. See each part's model in words under Models › Library.
• Tick Not electronic on a part neither handbook has (an airframe). It's left out of the total, and the result lists what was left out.
• For a part the handbook doesn't have, such as a battery, choose Give a rate of your own instead in the part-type search and type its failure rate per million hours, from the maker's data or your own records. The line is marked user-given everywhere it shows.
• A line with neither stays in the list with no part type: it's shown and left out of the total, and the result says so. Don't give it the nearest electronic part type just to get a number.

The sample prediction, the sample quadcopter's electronics and motors, shows a finished one with lines in assemblies: microcircuits on the handbook's generic rates, the motors on NSWC-11 with their conditions filled in, and the propellers left out as not electronic.

Read the result honestly

Only three of the handbook's factors are applied: λb, πE and πQ (or λg and πQ for a parts-count type, whose generic rate holds the rest at the handbook's defaults). The handbook's full models also have temperature, electrical stress, construction and power factors, and the result names which ones it left out for your parts. Real rates depend on them: temperature and stress alone can move a part's rate several times over. So read a parts count as it's meant to be read:

• For comparing designs early (does the second design have half the failure rate of the first?) and for finding what to improve (the Pareto), it's useful.
• As a measure of how your equipment will actually fail, it isn't. The handbook's models date from 1995 and are known to be pessimistic for modern parts; a prediction is often several times the failure rate seen in service.

The moment you have failures of your own, fit a model to them instead (fit a non-repairable model): data from your equipment beats a handbook every time. Saved predictions keep the result they were saved with; Edit computes it again from the parts list.