Reliability engineering

Production and Industrial Engineering (B.E./B.Tech) · Quality and reliability · free study notes for exams, placements and interviews

In short

Reliability engineering is the quantitative study of how long an item performs its intended function without failure, under stated conditions, for a stated time - and how to design, test and maintain it to raise that probability. It converts failure data into numbers a designer can use: failure rate, MTBF, reliability at mission time, availability of a plant. It is examined in GATE (PI and ME), in university papers on Quality and Reliability, and in operations-research and maintenance courses. In practice it drives warranty design at Maruti Suzuki and Bajaj, plant availability at NTPC and Indian Oil, mission assurance at ISRO and DRDO, and rolling-stock reliability at Indian Railways.

Explained

Definitions and the basic functions

Reliability R(t): probability that an item performs its required function, under given conditions, for the interval 0 to t. Always stated with a time/mission and an environment - "0.98 for 5,000 h at 40 °C" - never as a bare number.
Failure: termination of the ability to perform a required function. Fault is the state after failure; defect is a non-conformity that may or may not cause failure.
Unreliability (cumulative failure probability) F(t) = 1 − R(t); failure density f(t) = dF/dt.
Hazard (instantaneous failure) rate h(t) or λ(t) = f(t)/R(t): the failure rate of the units that have *survived* up to t. Units: failures per hour; electronics use FIT = 1 failure per 10⁹ component-hours.
MTTF - mean time to failure, for non-repairable items (a bearing, an LED, a solid rocket motor). MTBF - mean time between failures, for repairable items. MTTR - mean time to repair (a maintainability measure). MTBF = MTTF + MTTR strictly, but MTTR is usually negligible beside MTTF.
Maintainability: probability that a failed item is restored within a stated time. Availability: probability it is in a working state when needed.
Bathtub curve: three phases of h(t) - infant mortality (decreasing, from manufacturing and assembly defects; removed by burn-in and screening), useful life (roughly constant, random failures; the only region where the exponential model is honest), wear-out (increasing, from fatigue, corrosion, erosion; handled by preventive replacement).

Distributions used, and the derivation you are asked for

From h(t) = f(t)/R(t) = −(1/R)(dR/dt), integrate: R(t) = exp(−∫₀ᵗ h(u) du). This single relation generates every model below.
Exponential (h = λ constant): R(t) = e^−λt, MTTF = 1/λ. It is memoryless - a used unit is as good as new, so preventive replacement is useless for a purely exponential item.
Weibull (shape β, scale/characteristic life η, optional location γ): R(t) = exp[−((t−γ)/η)^β]. β < 1 → infant mortality, β = 1 → exponential, β ≈ 3.5 → nearly normal wear-out. Fitting β from a Weibull probability plot (or median ranks and least squares) tells you *which* part of the bathtub you are in - the most useful single technique in field-failure analysis.
Normal for wear-out dominated dimensions; lognormal for repair times and for fatigue/crack growth lives.
Poisson for the *number* of failures in time t when λ is constant: P(n) = (λt)ⁿ e^−λt/n! - the basis of spares provisioning.
MTTF = ∫₀^∞ R(t) dt for any distribution.

System reliability: configurations and redundancy

Reliability block diagram (RBD): success paths, not physical layout. Evaluated by series/parallel reduction, path-tracing, decomposition (conditional probability on a keystone element) or Monte Carlo.
Series: all must work. R_s = ΠR_i; for exponentials λ_s = Σλ_i, so the system MTBF is always less than the worst component's. Adding parts in series always lowers reliability.
Active parallel (redundant): any one suffices. R_p = 1 − Π(1 − R_i). Two identical exponential units: MTTF = 1.5/λ, not 2/λ.
Standby (sequential) redundancy with perfect switching and identical units: R(t) = e^−λt(1 + λt), MTTF = 2/λ - better than active parallel, but the switch/sensing reliability multiplies in and is usually the weak link.
k-out-of-n: R = Σ_{i=k}^{n} C(n,i) Rⁱ (1−R)ⁿ⁻ⁱ. Used in 2-out-of-3 voting logic (safety instrumented systems, triple-redundant avionics).
Levels of redundancy: component-level redundancy beats unit-level redundancy for the same part count; diversity (different design or vendor) protects against common-cause failure, which plain duplication does not.
Common-cause failure (shared power, shared cooling, same bad batch, one maintenance error) is what makes real redundant systems far worse than the formula predicts - β-factor models account for it.

Availability and maintainability

Inherent availability A_i = MTBF/(MTBF + MTTR) - corrective maintenance only. Achieved availability adds preventive maintenance; operational availability A_o adds logistic and administrative delay (spares waiting, permits, shift change) and is the number a plant manager actually sees.
Steady-state availability of a repairable unit = MTBF/(MTBF+MTTR) = μ/(λ+μ), where μ = 1/MTTR is the repair rate; derived from a two-state Markov model.
Availability rises faster by cutting MTTR (modularity, diagnostics, access, spares, trained crew) than by chasing MTBF, once MTBF is already large.
Design for maintainability: line-replaceable units, standard fasteners, built-in test, condition monitoring (vibration, oil analysis, thermography), fail-safe and fault-tolerant design.
Maintenance strategies: breakdown / corrective, scheduled preventive (time- or usage-based), condition-based, and reliability-centred maintenance (RCM), which selects the strategy per failure mode from its consequence - safety, environment, operational, hidden.

Reliability engineering through the life cycle

Specify: mission profile, environment, reliability/availability target, acceptable consequences.
Apportion (allocate): split the system target among subsystems - equal apportionment (each of n series blocks gets R^1/n), AGREE method, or weighting by complexity and criticality.
Predict: part-count or part-stress prediction from failure-rate handbooks (MIL-HDBK-217F, Telcordia SR-332, IEC 61709 for electronics) with environment and temperature factors. Predictions are for *comparison*, not promises.
Analyse: FMEA/FMECA (bottom-up, mode by mode), FTA (top-down, Boolean, to minimal cut sets), ETA, HAZOP, worst-case and derating analysis, sneak-circuit and thermal analysis.
Design actions: derating, simplification, redundancy, robust (Taguchi) design, screening of parts, protective devices.
Test: reliability development/growth testing, HALT (highly accelerated life test to find weak links), ALT (accelerated life test with a physical acceleration model - Arrhenius for temperature, inverse power law for voltage/stress, Coffin-Manson for thermal cycling), reliability demonstration/qualification test, ESS/burn-in, HASS in production.
Track in service: FRACAS (failure reporting, analysis and corrective action), warranty and field data, reliability growth monitoring - Duane model cumulative MTBF = T^α/k plotted log-log, growth slope α typically 0.3-0.6; Crow-AMSAA is its statistical form.
FMEA: rank by RPN = Severity × Occurrence × Detection, each 1-10. The AIAG-VDA FMEA handbook (2019, used by Indian automotive suppliers) replaced RPN ranking with Action Priority (High/Medium/Low) tables, because equal RPNs can hide very different severities - say so if asked about RPN's limitation.
Design review, change control and configuration management are part of reliability work: an untracked change invalidates all earlier analysis.

Statistics of reliability data, and Indian standards

Censoring: data are usually incomplete - right-censored/suspended units that had not failed when the test stopped. Ignoring suspensions understates life badly; use median-rank regression or maximum likelihood.
Point estimate of MTBF from a test = total unit-hours T / number of failures r.
Confidence interval (exponential): time-terminated test lower limit MTBF_L = 2T/χ²_{α,(2r+2)}; failure-terminated uses 2r degrees of freedom.
Reliability demonstration: a zero-failure test needs total test time n·t = MTBF_required × ln[1/(1−C)] = −MTBF·ln(1−C); for 90% confidence, about 2.3 × MTBF of total unit-hours with no failure.
Standards and practice in India: BIS adopts the IEC dependability series as IS/IEC - IEC 60300 (dependability management), IEC 60812 (FMEA/FMECA), IEC 61025 (fault tree analysis), IEC 61078 (reliability block diagrams), IEC 61709 (component failure-rate reference conditions), IEC 61508 and IEC 61511 (functional safety, SIL - safety integrity level defined by probability of failure on demand), ISO 55000 (asset management). Where an IS number has been revised or withdrawn, quote the subject rather than a number.
Sectoral practice: AERB requires probabilistic safety assessment for nuclear plants; CEA/CERC availability-based tariff makes plant availability a commercial number; TRAI quality-of-service norms impose network availability on telecom operators; the Occupational Safety, Health and Working Conditions Code, 2020, in force from 21 November 2025 in place of the Factories Act, governs safety of plant and equipment and pressure-vessel/lifting-tackle testing. Automotive reliability and durability validation follows AIS standards and homologation testing at ARAI/ICAT.

Key points

Reliability - probability of no failure for a stated time, function and environment; meaningless without all three.
Failure rate λ - failures per unit time of survivors; FIT = 10⁻⁹ /h.
MTBF ≠ guaranteed life - with constant λ, only ~36.8% of units survive to MTBF (e⁻¹).
Bathtub curve - decreasing (infant mortality), constant (random), increasing (wear-out) hazard.
Exponential model - constant λ, memoryless, R = e^−λt, MTTF = 1/λ; valid only in useful life.
Weibull β - the diagnostic: β<1 infant mortality, β=1 random, β>1 wear-out.
R(t) = exp(−∫h dt) - the master relation linking hazard and reliability.
Series - λ adds, reliability multiplies, system is weaker than its weakest part.
Parallel - unreliabilities multiply; two units give MTTF = 1.5/λ.
Standby - R = e^−λt(1+λt), MTTF = 2/λ, but limited by switch reliability.
Common-cause failure - the real limit on redundancy; defeated by diversity and separation, not duplication.
Availability = MTBF/(MTBF+MTTR); cutting MTTR is often the cheaper lever.
FMEA - bottom-up, mode by mode, RPN = S×O×D (AIAG-VDA now prefers Action Priority).
FTA - top-down Boolean, output is minimal cut sets and top-event probability.
Burn-in removes infant mortality; preventive replacement addresses wear-out; neither helps a purely exponential item.
Accelerated testing needs a physical model (Arrhenius, inverse power law) and an acceleration factor to extrapolate.
Censored data must be included as suspensions; dropping them overstates failure.
Reliability growth (Duane/Crow-AMSAA) measures whether test-fix-test is actually improving the design.

Formulas and examples

R(t) = 1 − F(t); f(t) = dF/dt; h(t) = f(t)/R(t); R(t) = exp(−∫₀ᵗ h(u)du) [t in hours, h in failures/hour]
Exponential: R(t) = e^−λt; MTTF = 1/λ; λ = r/T (r = failures, T = total unit-hours)
Weibull: R(t) = exp[−(t/η)^β]; h(t) = (β/η)(t/η)^β−1; MTTF = η·Γ(1 + 1/β); B10 life = η(0.10536)^1/β
MTTF = ∫₀^∞ R(t) dt
Series: R_s = ΠR_i; λ_s = Σλ_i; MTBF_s = 1/Σλ_i
Active parallel (n identical): R_p = 1 − (1−R)ⁿ; MTTF (n=2) = 1.5/λ
Standby (2 units, perfect switch): R(t) = e^−λt(1+λt); MTTF = 2/λ
k-out-of-n (identical): R = Σ_{i=k}^{n} C(n,i)Rⁱ(1−R)ⁿ⁻ⁱ
Availability A = MTBF/(MTBF+MTTR) = μ/(λ+μ); Unavailability × 8,760 = downtime hours/year
Poisson spares: P(n failures) = (λt)ⁿe^−λt/n!
Equal apportionment: R_i = R_system^1/n (n series blocks)
MTBF lower confidence limit (time-terminated) = 2T/χ²_{α,2r+2}
Arrhenius acceleration factor AF = exp[(Ea/k)(1/T_use − 1/T_test)], T in kelvin, k = 8.617×10⁻⁵ eV/K
RPN = Severity × Occurrence × Detection (1-10 each, 1-1000)

Example 1 - exponential reliability and the MTBF trap. A pump has a constant failure rate λ = 2×10⁻⁴ /h.

MTBF = 1/λ = 5,000 h.
R(1,000 h) = e^−0.2 = 0.8187 (81.9% survive a 1,000-hour campaign).
R(5,000 h) = e⁻¹ = 0.3679 - only 36.8% reach the MTBF.
Time for 90% reliability: t = −ln(0.90)/λ = 0.10536/2×10⁻⁴ = 527 h. So an "MTBF 5,000 h" pump needs servicing roughly every 500 h for a 90% success target.

Example 2 - mixed system. A press line: hydraulic power pack A (R = 0.90) in series with two redundant control PLCs, each R = 0.80, and a feeder F (R = 0.95), all for a 2,000-hour mission.

PLC pair (active parallel): 1 − (1−0.80)² = 1 − 0.04 = 0.96.
System: 0.90 × 0.96 × 0.95 = 0.8208.
Replace the PLC pair by a single PLC: 0.90 × 0.80 × 0.95 = 0.6840. Redundancy on one block raised system reliability by 13.7 percentage points.
If instead the two PLCs are in cold standby with a changeover relay of reliability 0.98: pair ≈ 0.80 + 0.98 × (1−0.80) × 0.80 ≈ 0.957 - slightly *worse* than active parallel here because the relay can fail.

Example 3 - Weibull wear-out of a bearing. Field data fit β = 2.0, η = 1,000 h.

R(500) = exp[−(500/1000)²] = exp(−0.25) = 0.7788.
h(500) = (2/1000)(500/1000)¹ = 1.0×10⁻³ /h; h(1,000) = (2/1000)(1) = 2.0×10⁻³ /h - the hazard has doubled, confirming wear-out.
MTTF = 1,000 × Γ(1.5) = 1,000 × 0.8862 = 886 h (note: less than η).
B10 life = 1,000 × (0.10536)^1/2 = 1,000 × 0.3246 = 325 h - the number a bearing catalogue (L10) would quote.
Because β > 1, preventive replacement at about 325-500 h is justified; for β = 1 it would not be.

Example 4 - availability and the MTTR lever. A CNC machine: MTBF 400 h, MTTR 8 h.

A = 400/408 = 0.9804 → downtime = 8,760 × 0.0196 = 172 h/year.
Halve MTTR to 4 h (spares kit + onboard diagnostics): A = 400/404 = 0.9901 → downtime 87 h/year.
To reach the same availability by MTBF alone you would need MTBF ≈ 800 h - double the design reliability. The maintainability route is cheaper.

Example 5 - MTBF estimate, confidence limit and spares. Twenty identical gearboxes run 500 h each; 4 fail (failed units replaced, test stopped at 500 h).

T = 20 × 500 = 10,000 unit-hours, r = 4 → MTBF = 10,000/4 = 2,500 h, λ = 4×10⁻⁴ /h.
90% lower confidence limit: χ²_{0.10, 2r+2=10} = 15.987 → MTBF_L = 2×10,000/15.987 = 1,251 h. Report "MTBF 2,500 h, ≥1,251 h at 90% confidence".
Spares for one gearbox over 1,250 h: λt = 0.5. P(0 failures) = e^−0.5 = 0.6065; with 1 spare, P(≤1) = e^−0.5(1+0.5) = 0.9098; with 2 spares, P(≤2) = e^−0.5(1+0.5+0.125) = 0.9856. Two spares meet a 98% service level.

Configuration (2 identical units, λ each)	R at t	MTTF	Comment
Series	e^−2λt	0.5/λ	Worse than one unit
Single unit	e^−λt	1/λ	Baseline
Active parallel	1 − (1−e^−λt)²	1.5/λ	Both energised, shared load risk
Cold standby (perfect switch)	e^−λt(1+λt)	2/λ	Best, if the switch is reliable

β in Weibull	Hazard trend	Phase	Correct action
β < 1	Decreasing	Infant mortality	Burn-in, screening, process control
β = 1	Constant	Useful life	Redundancy, derating; no PM benefit
1 < β < 3	Increasing	Early wear-out	Condition monitoring
β ≥ 3	Sharply increasing	Wear-out	Scheduled replacement at B10/B5

Where it is used

Space and defence: ISRO's launch-vehicle and satellite reliability budgets, redundant avionics and triple-redundant control; DRDO qualification testing and FMECA on missile and radar programmes.
Automotive: warranty-driven reliability at Maruti Suzuki, Tata Motors, Mahindra, Bajaj and TVS - DFMEA/PFMEA under AIAG-VDA and IATF 16949, B10 life targets for bearings and clutches, accelerated durability on ARAI/ICAT road-load cycles.
Power and process plants: NTPC, Tata Power, Reliance and IOCL track plant availability, forced-outage rate and equipment MTBF; CERC availability-based tariff makes availability a revenue term; AERB requires PSA for NPCIL nuclear stations.
Railways and transit: Indian Railways and Vande Bharat rolling-stock MTBF per million km, RDSO reliability specifications, metro signalling with SIL-4 fail-safe interlocking under IEC 61508/EN 50129.
Electronics, telecom and IT: FIT-based prediction for PCBs, data-centre and network availability ("five nines" = 5.3 min downtime/year), TRAI QoS obligations, HALT/HASS in EMS plants.
Job roles: reliability engineer, maintenance/asset-management engineer, quality and NPD engineer (FMEA owner), test and validation engineer, safety/functional-safety engineer (CFSE), plant reliability analyst using CMMS/SAP PM data.

Interview questions

Q: If a machine's MTBF is 1,000 hours, what is the chance it runs 1,000 hours without failure? A: With a constant failure rate, R = e⁻¹ ≈ 36.8% - MTBF is a mean, not a guaranteed life.
Q: Difference between MTTF and MTBF? A: MTTF is the mean life of a non-repairable item; MTBF is the mean operating time between failures of a repairable item, which is restored and returned to service.
Q: What does the Weibull shape parameter tell you? A: β < 1 means infant mortality (fix the process or burn in), β = 1 means random failures, β > 1 means wear-out (scheduled replacement makes sense).
Q: Why does adding components in series always reduce reliability? A: Reliabilities multiply and each is below 1, and failure rates add - so the system MTBF is lower than the worst component's.
Q: Which is better, active parallel or cold standby? A: Cold standby gives a higher MTTF (2/λ against 1.5/λ) because the spare does not age, but it depends on a sensing-and-switching device whose own failure can wipe out the gain.
Q: Two redundant pumps on the same power feed - is reliability really 1−(1−R)²? A: No. The shared feed is a common-cause failure that makes the pair effectively a series element with the supply; redundancy must be separated and preferably diverse.
Q: How do you raise availability from 98% to 99% most cheaply? A: Usually by halving MTTR - modular replacement, diagnostics, spares on site and trained crew - rather than doubling MTBF through redesign.
Q: What is the difference between FMEA and FTA? A: FMEA is inductive and bottom-up, listing each failure mode and its effects; FTA is deductive and top-down, starting from an undesired top event and finding Boolean minimal cut sets.
Q: Criticism of RPN? A: Equal RPNs can hide very different severities, the scales are ordinal so multiplying them is not strictly valid, and many RPN values are unreachable - hence the AIAG-VDA Action Priority tables.
Q: What is burn-in, and when is it pointless? A: Operating units before shipment to weed out infant mortality; it is pointless (and wastes life) when the hazard rate is constant or increasing, i.e. β ≥ 1.
Q: How do you use accelerated life testing to predict field life? A: Stress the item above use conditions, fit a life model, and convert with an acceleration factor from a physical law - Arrhenius for temperature, inverse power law for voltage, Coffin-Manson for thermal cycling - valid only if the failure mode is unchanged.
Q: What is SIL? A: Safety integrity level under IEC 61508/61511 - a band of required risk reduction expressed as probability of failure on demand, achieved by architecture (e.g. 2-out-of-3), proof-test interval and diagnostic coverage.
Q: You have censored data - 50 units tested, 6 failed, 44 still running. How do you estimate life? A: Treat the 44 as suspensions in median-rank regression or maximum-likelihood Weibull fitting; never discard them, or you will badly understate life.

Common mistakes

Treating MTBF as a guaranteed or typical life, or as "the machine will fail at 5,000 hours".
Applying the exponential formula to wear-out items such as bearings, belts and tyres, where β > 1 and the memoryless assumption is false.
Assuming MTTF of two parallel units is 2/λ (it is 1.5/λ) or forgetting the switch reliability in standby systems.
Computing redundancy gains while ignoring common-cause failures, shared supplies and the same faulty batch.
Dropping suspended (censored) units from the data, or counting calendar hours instead of operating unit-hours in λ = r/T.
Mixing units and time bases - FIT with failures/hour, hours with cycles or kilometres - and forgetting that λ must match the mission's clock.
Reporting a point MTBF with no confidence interval, or extrapolating an accelerated test without checking that the failure mode is the same as in service.
Confusing reliability with quality at delivery: a part can pass every incoming inspection and still have a short life.