Reliability Probability and Statistical Foundations

Introduction and context

Reliability engineering applies probability and statistics to the question: how likely is an item, system, or process to perform its required function for a specified time, under stated conditions? Mechanical engineers use these ideas to select components, plan preventive maintenance, compare designs, estimate warranty exposure, and make decisions when failure data are incomplete.

A reliability statement is meaningful only when its function, operating environment, mission time, and failure criterion are defined. For example, “the pump is reliable” is incomplete. A better statement is: “The pump has at least 0.95 probability of delivering 20 L/s continuously for 1,000 operating hours at 25 °C, provided suction pressure remains within the specified range.”

This chapter develops the probability models needed for reliability analysis, including survival and hazard functions, mean time measures, common failure-rate models, censored observations, field data, and uncertainty. The emphasis is not merely on calculation, but on interpreting evidence responsibly.

Learning objectives

By the end of this chapter, the reader should be able to:

1. Define reliability, unreliability, failure time, survival probability, and hazard rate.
2. Relate the probability density, cumulative distribution, reliability, and hazard functions.
3. Calculate and interpret MTTF, MTBF, failure rate, and availability-related measures.
4. Apply exponential, Weibull, and normal or lognormal models appropriately.
5. Analyse complete, right-censored, left-censored, and interval-censored data.
6. Estimate reliability from field data while recognising sampling, measurement, and model uncertainty.
7. Communicate reliability results with appropriate assumptions, confidence, and data-quality qualifications.

1. Fundamental probability concepts

Let T denote the non-negative random variable representing time to failure. Its units may be hours, cycles, kilometres, or another appropriate exposure measure. The analysis assumes that the population of items and the operating conditions have been defined.

The cumulative distribution function (CDF) is

where F(t) is the probability that failure has occurred by time t. The reliability or survival function is

For continuous T, the probability density function is

The density f(t) is not itself a probability at a particular time; probability is obtained by integrating it over an interval. Thus,

The functions satisfy F(0) = 0 and R(0) = 1 when all items are new at t = 0, and F(t) + R(t) = 1. A valid continuous lifetime model has f(t) ≥ 0 and ∫[0 to ∞] f(t)dt = 1.

Conditional survival and the hazard function

Reliability conditional on survival to time u is

The instantaneous hazard rate is defined as

For a continuous distribution,

Hazard has units of inverse time, such as failures per hour. It is a conditional instantaneous risk, not the probability of failure over a whole mission. For a small interval ∆t, the conditional failure probability is approximately h(t)∆t when h(t)∆t is small.

The cumulative hazard is

and therefore

These relationships require a continuous lifetime model and a well-defined time origin. They are especially useful because a model can be specified through either its hazard or survival function.

2. Mean time measures

The mean time to failure (MTTF) is the expected lifetime of a non-repairable item:

The second equality follows by integration by parts and applies when the expected lifetime is finite. MTTF is appropriate for bearings, fuses, sealed electronic modules, or complete machines that are discarded after failure.

For a repairable population, mean time between failures (MTBF) is commonly calculated as

This empirical measure is useful only when the numerator and denominator are clearly defined. In a long-run renewal process with complete restoration and stable conditions, the average cycle length between successive failures and repairs may be interpreted as a mean inter-failure time. However, MTBF is not automatically the useful life of a non-repairable product, nor does a high MTBF identify the cause or timing of failures.

Mean time to repair (MTTR) is commonly

A simple steady-state availability approximation is

This assumes alternating operating and repair periods, stable rates, and negligible logistics or waiting distinctions. If administrative delay, spares delay, and active repair time differ, they should be reported separately rather than hidden inside MTTR.

Worked example: empirical measures

A compressor operates for 18,000 h and experiences 6 corrective failures. The recorded corrective repair time is 150 h. The empirical values are

This result describes the observed operating and repair record under the stated boundary. It does not prove that every future compressor has a 3,000 h MTBF, especially if operating severity, maintenance quality, or failure detection changes.

3. Probability models used in reliability

3.1 Exponential model

The exponential model assumes a constant hazard rate λ, with units such as h⁻¹:

The model is memoryless: conditional on surviving to u, the remaining lifetime distribution is the same as that of a new item. This can be reasonable for random electronic failures in a stable period, but is usually inappropriate when wear-out is dominant.

If an item has MTTF = 10,000 h and the exponential assumption is accepted, λ = 0.0001 h⁻¹. Reliability at 2,000 h is

The calculation is model-dependent; MTTF alone does not determine reliability unless the distributional form is also specified.

3.2 Weibull model

The two-parameter Weibull distribution is widely used for mechanical components:

Here β > 0 is the shape parameter and η > 0 is the scale parameter, with the same time units as t. At t = η, R(η) = e⁻¹ ≈ 0.3679; η is therefore a characteristic life, not generally the MTTF.

The hazard interpretation is central:

For the Weibull model,

where Γ is the gamma function. This formula assumes a complete two-parameter Weibull population beginning at t = 0. If a location parameter is used, the expression must be modified.

3.3 Normal and lognormal models

A normal lifetime model can assign non-zero probability to negative time and is therefore unsuitable unless the probability of negative values is negligible or the variable being modelled is not lifetime itself. The lognormal model is often more appropriate for positive lifetimes affected by multiplicative factors. If ln(T) is normally distributed with parameters μ and σ², then

where Φ is the standard normal CDF. Lognormal models are useful for some fatigue, degradation, and repair-time data, but model selection should be based on physical reasoning and diagnostic evidence, not visual fit alone.

4. Series, parallel, and system reliability

For statistically independent components in a series system, the system operates only if every component operates:

For two independent components in parallel, the system operates if at least one operates:

These equations require a correct system success definition and independence. Common-cause failures, shared utilities, environmental exposure, and maintenance interactions can invalidate the simple formulas.

For a series system of independent exponential components with rates λ_i, the system hazard is constant and equals

This is a special case, not a general rule for all distributions.

5. Estimating reliability from data

Complete data

Complete data are observations for which the failure time of every item is known. If n independent items fail at times t₁, t₂, …, tₙ, the empirical estimate of reliability at mission time t can be formed from the proportion surviving beyond t:

This non-parametric estimate is simple but changes in steps and can be imprecise for small samples. A parametric estimate fits a selected distribution and estimates its parameters, commonly by maximum likelihood.

For complete exponential data, the maximum-likelihood estimate of λ is

This assumes independent lifetimes, a constant hazard, and a common exposure origin.

Right-censored data

In reliability testing, some units have not failed when the test ends. Their exact failure time is unknown; we only know it exceeds the censoring time. Such observations are right-censored, not failures at the test-ending time.

Let δ_i = 1 if unit i fails and δ_i = 0 if it is censored, and let x_i be the observed time (failure or censoring time). For an exponential model with right censoring, the likelihood is

which simplifies to

where r = ∑δ_i is the number of failures. Thus,

All observed time contributes exposure, but only actual failures contribute to r. Treating censored units as failures biases reliability downward.

Other censoring types include left censoring, where failure occurred before observation began, and interval censoring, where failure is known only to have occurred between two inspections. Inspection-based maintenance data are often interval-censored and should not automatically be converted into exact failure times.

Kaplan–Meier estimate

The Kaplan–Meier estimator is a non-parametric approach that accommodates right-censored observations. At each distinct failure time t_j, let d_j be the number of failures and n_j the number at risk immediately before that time:

It estimates the survival curve without imposing exponential or Weibull shape assumptions. Confidence intervals are important, particularly in the tail where few units remain at risk.

Field data and data quality

Field data may include operating hours, starts, loads, temperatures, maintenance actions, removed-part status, and failure codes. Before modelling, engineers should reconcile asset identity, time zones, meter resets, duplicate records, missing periods, and inconsistent definitions of “failure.” Exposure should be measured in a physically meaningful scale. A pump may be better modelled against running hours, while a fatigue component may be better modelled against cycles or load spectrum.

A useful data-quality checklist asks:

• Is the failure definition linked to loss of required function?
• Are operating and standby periods distinguished?
• Are preventive replacements recorded separately from failures?
• Are failures detected immediately or only at inspection?
• Are repairs minimal, imperfect, or complete?
• Are multiple failure modes mixed together?
• Have changes in design, supplier, software, or maintenance policy been identified?

6. Uncertainty and statistical interpretation

A reliability estimate is not a fact without uncertainty. Sources include sampling variation, censoring, parameter uncertainty, model uncertainty, measurement error, and changing operating conditions.

For a binomial estimate based on n independent identical trials, the approximate standard error of a proportion p̂ is

This approximation is less reliable when n is small or p̂ is near 0 or 1; exact or Wilson confidence intervals are preferable in those circumstances. Parametric confidence intervals may be obtained from likelihood methods, while bootstrap methods can be useful when assumptions are difficult to justify.

Confidence intervals describe uncertainty in a statistical procedure under repeated sampling. They do not mean that a fixed parameter has a stated probability of lying in one particular calculated interval. Bayesian credible intervals have a different interpretation and require a prior distribution.

Reliability demonstration tests must also distinguish producer and consumer risks: a test can fail to detect an inadequate design or reject an adequate one. Zero failures is not proof of perfect reliability. For an exponential model with zero failures in total accumulated test time T, an approximate one-sided 90% upper confidence bound on λ is

The corresponding lower confidence bound on MTTF is T/2.303, under the stated exponential assumptions. This result illustrates why large exposure is needed to substantiate high reliability.

7. Failure-rate models and the bathtub curve

The bathtub curve is a conceptual combination of three regimes: decreasing hazard during early defects, approximately constant hazard during useful life, and increasing hazard during wear-out. It is not a universal empirical law and should not be imposed on every product.

Failure modes may overlap, and different subpopulations may have different hazards. Mixture populations can produce a declining aggregate hazard even when individual subgroups do not. Consequently, failure-mode separation, stratification by operating conditions, and physical failure analysis are often more informative than fitting one curve to all events.

Preventive replacement is justified only when it reduces risk or lifecycle cost. Replacing all items at a fixed age does not necessarily improve reliability when failures are random (β ≈ 1), and premature replacement can introduce installation defects.

8. Illustrative case: cooling-fan field reliability

Background

A manufacturer receives field reports for 40 industrial cooling fans. The engineering requirement is reliability at 5,000 operating hours. The fans are monitored for up to 5,000 h; some remain operating at the end of observation.

Problem

The raw records contain 14 observed failures and 26 right-censored observations. Total exposure, including both failed and surviving fans, is 172,000 h. Management initially proposes calculating failure rate as 14 divided by the exposure and treating every surviving fan as a 5,000 h success without further analysis.

Analysis

For an initial exponential screening model,

The estimated reliability at 5,000 h is

This estimate is substantially below the company target of 0.90. However, the engineer checks the data before accepting the conclusion. Six failures occurred in a dusty installation category, five involved a revised bearing supplier, and three were electrical failures. The exposure is not evenly distributed across categories. Some “failures” were actually preventive removals, and inspection records show that several units may have failed between inspections.

Approach and solution

The team separates true functional failures from preventive removals, analyses failure modes independently, records inspection intervals as interval-censored observations, and compares exposure by installation category. A Kaplan–Meier curve is used for descriptive analysis, followed by Weibull modelling for the bearing-related subgroup. Environmental controls and a supplier corrective action are introduced. The reliability target is then reassessed using a prospective verification test rather than the original mixed field population.

Outcome

The original exponential estimate remains a useful warning signal, but not a final design conclusion. Stratification identifies dust ingress and the supplier change as actionable contributors. The revised decision is based on a defined failure criterion, verified exposure, an explicit confidence requirement, and a test population representative of intended service.

Learning points

1. Censored observations contribute information and must not be discarded or treated as failures.
2. A single aggregate failure rate can conceal important subpopulations.
3. Model output cannot compensate for ambiguous failure definitions.
4. Reliability improvement should be connected to physical failure mechanisms and controlled verification.

9. Professional practice, governance, and contemporary tools

Reliability data influence safety, warranty, maintenance expenditure, and public risk. Engineers must not suppress inconvenient failures, redefine failure to improve a reported metric, or present a point estimate without its exposure and uncertainty. Traceability is essential: each result should identify the dataset version, inclusion rules, model, assumptions, and confidence method.

Digital maintenance systems, sensors, and fleet telematics increase data volume but not automatically data validity. Machine-learning models can detect anomalies or estimate remaining useful life, but they require representative labelled data, drift monitoring, explainable decision thresholds, and human oversight. A prediction is not a reliability estimate unless its probability calibration and validation performance have been demonstrated. Privacy, cybersecurity, access control, and retention policies also apply to connected equipment data.

Best practice is to combine statistical analysis with engineering judgement: define function and failure first; measure exposure; preserve censored information; inspect data quality; select models using physics and diagnostics; quantify uncertainty; validate predictions; and revise the model when design or operating conditions change.

Chapter Summary

Reliability is the probability that an item performs its required function for a specified time and condition. The survival function R(t), failure distribution F(t), density f(t), hazard h(t), and cumulative hazard H(t) are mathematically related but have different interpretations. MTTF applies primarily to non-repairable items, whereas MTBF is an empirical repairable-system measure whose meaning depends on boundaries and data definitions.

The exponential model assumes constant hazard; the Weibull model permits decreasing, constant, or increasing hazard through its shape parameter. Series and parallel formulas require appropriate system definitions and independence. Complete, right-censored, left-censored, and interval-censored data must be handled according to what is actually known. Kaplan–Meier estimation provides a distribution-free survival estimate, while parametric models offer compact extrapolation only when assumptions are credible.

Reliability results always contain uncertainty. Sampling variation, censoring, model choice, measurement quality, mixture populations, and changing environments must be considered. Sound practice combines statistical evidence with failure physics, transparent governance, and appropriately validated digital tools.

Key Terms

Reliability: Probability that an item performs its required function for a stated time and condition.

Failure time: Random time at which an item ceases to meet its specified function.

Survival function: Probability that failure time exceeds t, R(t) = P(T > t).

Cumulative distribution function: Probability that failure occurs by t, F(t) = P(T ≤ t).

Probability density: Rate of accumulation of failure probability in continuous time.

Hazard rate: Instantaneous conditional failure intensity among items surviving to time t.

Cumulative hazard: Integral of hazard over time.

MTTF: Expected time to failure for a non-repairable item.

MTBF: Operating time divided by observed number of failures for a repairable context.

MTTR: Average corrective-maintenance time per repair.

Availability: Proportion of time an item is capable of performing its function.

Exponential distribution: Lifetime model with constant hazard.

Weibull distribution: Flexible lifetime model controlled by shape and scale parameters.

Right censoring: Knowing only that failure time exceeds an observed time.

Interval censoring: Knowing that failure occurred between two observation times.

Kaplan–Meier estimator: Non-parametric estimate of a survival curve with censoring.

Complete data: Observations for which failure times are known.

Model uncertainty: Uncertainty arising from selecting an imperfect probability model.

Common-cause failure: Failure mechanism affecting multiple supposedly independent components.

Exposure: Time, cycles, distance, or other measured opportunity for failure.

Review Questions

5 Short Answer Questions

1. Distinguish reliability R(t) from unreliability F(t).
2. Why is hazard rate not the same as the probability of failure over an entire mission?
3. State one appropriate use of MTTF and one limitation of MTBF.
4. Interpret Weibull shape parameters β < 1, β = 1, and β > 1.
5. What information does a right-censored observation provide?

5 Long Answer Questions

1. Derive the relationship between the survival function and cumulative hazard, and explain the assumptions involved.
2. Compare exponential, Weibull, and lognormal lifetime models in terms of assumptions and engineering applications.
3. Explain how complete, censored, and interval-censored data should be treated in a field reliability study.
4. Discuss why a bathtub curve should be regarded as a conceptual model rather than a universal law.
5. Explain the differences among statistical uncertainty, model uncertainty, and data-quality limitations.

3 Analytical / Application Questions

1. Ten identical pumps accumulate 32,000 operating hours and experience four failures. Assuming a constant hazard, estimate λ and reliability at 1,000 h. State the limitations of the calculation.
2. A two-component series system has independent exponential component rates of 2 × 10⁻⁵ h⁻¹ and 3 × 10⁻⁵ h⁻¹. Calculate system reliability at 1,000 h and explain whether independence is realistic in the proposed application.
3. A field dataset contains many preventive removals, inspection-detected failures, and a supplier change midway through observation. Design an analysis plan that preserves censoring information, separates populations, and reports uncertainty responsibly.

Multiple Choice Questions with Answers

Question 1

Which expression defines reliability for a continuous lifetime T?

A. P(T ≤ t)

B. P(T > t)

C. f(t)/t

D. ∫[0 to t] h(u)du

Correct Answer: B. P(T > t)

Question 2

A constant hazard rate is the defining assumption of which model?

A. Normal

B. Weibull with β > 1

C. Exponential

D. Lognormal

Correct Answer: C. Exponential

Question 3

A component has an exponential failure rate of 0.0001 h⁻¹. Its reliability at 2,000 h is approximately:

A. 0.1813

B. 0.5000

C. 0.8187

D. 0.9998

Correct Answer: C. 0.8187

Question 4

For a Weibull distribution, what does β > 1 generally indicate?

A. Decreasing hazard

B. Constant hazard

C. Increasing hazard

D. Zero hazard

Correct Answer: C. Increasing hazard

Question 5

A unit operating at the end of a test without failing is usually recorded as:

A. An exact failure at test end

B. A right-censored observation

C. A left-censored observation

D. A duplicate failure

Correct Answer: B. A right-censored observation

Question 6

For independent components in series, system reliability is:

A. The sum of component reliabilities

B. The maximum component reliability

C. The product of component reliabilities

D. One minus the product of component reliabilities

Correct Answer: C. The product of component reliabilities

Question 7

A repairable machine operates for 12,000 h and has 4 failures. Its empirical MTBF is:

A. 48 h

B. 750 h

C. 3,000 h

D. 12,004 h

Correct Answer: C. 3,000 h

Question 8

Which method estimates a survival curve without imposing an exponential or Weibull distribution?

A. Kaplan–Meier estimation

B. Series reduction

C. MTTR calculation

D. Bathtub fitting only

Correct Answer: A. Kaplan–Meier estimation

Question 9

Under an exponential model with right censoring, the estimated failure rate is calculated as:

A. Number censored divided by number failed

B. Number failed divided by total observed exposure

C. Total exposure divided by number censored

D. Mean censoring time divided by total failures

Correct Answer: B. Number failed divided by total observed exposure

Question 10

Why should a reliability estimate normally be reported with uncertainty information?

A. Because reliability is always exactly normally distributed

B. Because point estimates are affected by sampling, model, and data limitations

C. Because uncertainty makes all models invalid

D. Because confidence intervals replace failure definitions

Correct Answer: B. Because point estimates are affected by sampling, model, and data limitations