2. Reliability Theory and Failure Modelling

Introduction and context

Maintenance decisions are ultimately decisions about uncertainty. A machine may operate for thousands of hours, fail shortly after commissioning, or deteriorate gradually until its performance becomes unacceptable. Reliability theory provides the mathematical language for describing these possibilities. It enables maintenance professionals to estimate the probability that an asset will perform its intended function, identify the nature and timing of failures, compare design alternatives, and determine suitable inspection, replacement, and spare-parts policies.

Reliability is not the same as quality, durability, availability, or maintainability. Quality concerns conformance at the time of manufacture; durability concerns useful life; maintainability concerns the ease and speed of restoration; availability combines operational readiness with failure and repair processes. Reliability focuses specifically on failure-free performance under stated conditions for a stated period. These distinctions are essential in production management because a highly reliable machine may still have poor availability if repairs are excessively long, while a rapidly repairable machine may remain available despite frequent failures.

This chapter develops the principal models used in reliability and maintenance management. It begins with probability concepts and reliability functions, then examines failure distributions, hazard rates, bathtub curves, life measures, repairable systems, system configurations, and Weibull analysis. The final sections apply these ideas to managerial decisions and an illustrative industrial case.

Learning objectives

After studying this chapter, the reader should be able to:

1. Define reliability, failure, maintainability, availability, MTTF, MTBF, and MTTR.

2. Derive and interpret reliability, unreliability, density, survival, and hazard functions.

3. Select suitable probability distributions for different failure mechanisms.

4. Explain the bathtub curve and distinguish failure-rate regions.

5. Calculate basic measures for non-repairable and repairable items.

6. Analyse series, parallel, and mixed system configurations.

7. Estimate and interpret Weibull parameters from failure data.

8. Use reliability results in maintenance, design, spare-parts, and operational decisions.

9. Recognise modelling assumptions, data limitations, and ethical risks in reliability analysis.

Reliability as a probabilistic performance concept

Let T denote the random variable representing time to failure of an item. Reliability at time t is the probability that the item performs its required function without failure up to t:

R(t) = P(T > t).

The complementary probability is unreliability or the cumulative distribution function:

F(t) = P(T ≤ t) = 1 − R(t).

If failure times have a probability density f(t), then:

F(t) = ∫₀ᵗ f(u)du, and R(t) = ∫ₜ∞ f(u)du.

The density describes the relative concentration of failures around a particular time. It should not be confused with the probability of failure at an exact instant, which is zero for a continuous distribution.

The hazard function, or instantaneous failure rate, is

h(t) = f(t)/R(t).

It represents the conditional rate of failure at time t, given survival to t. In practical terms, it answers the question: among items still operating at age t, how strongly are they exposed to failure at that age? A hazard rate is not generally a probability and may exceed one when measured per unit of time. The cumulative hazard is

H(t) = ∫₀ᵗ h(u)du,

and R(t) = exp[−H(t)].

These functions are mathematically linked, but they convey different managerial meanings. Reliability is useful for warranty and mission planning; the density helps identify when failures are concentrated; the hazard function supports preventive-maintenance reasoning.

Failure, fault, and failure modes

A fault is a physical or logical defect, condition, or abnormality. A failure occurs when the item can no longer perform a required function within specified limits. A component can therefore contain a fault without yet having failed. Failure criteria must be defined operationally. For a pump, complete stoppage is one criterion, but unacceptable flow, leakage, vibration, or energy consumption may also constitute functional failure.

Failure modes identify how failure occurs: bearing seizure, insulation breakdown, seal leakage, software lock-up, fatigue cracking, or loss of calibration. Failure effects describe the consequences at equipment, production, safety, environmental, and business levels. Failure mechanisms explain the physical process, such as fatigue, wear, corrosion, overheating, contamination, or random electronic breakdown. This distinction is important because a single equipment-level failure may have several modes, each requiring a different maintenance policy.

Reliability analysis commonly assumes a defined population, specified operating environment, stated mission time, and clear success criterion. Changing load, temperature, speed, duty cycle, maintenance quality, or operator practice changes the underlying failure distribution.

Probability models for failure behaviour

Exponential model

The exponential distribution assumes a constant hazard rate λ:

R(t) = e^(−λt),

F(t) = 1 − e^(−λt),

f(t) = λe^(−λt).

Its mean time to failure is

MTTF = 1/λ.

The model has the memoryless property: the probability of failure in the next interval does not depend on age, provided the item has survived to the beginning of that interval. This assumption can be reasonable for certain electronic components and random shocks, but it is usually inappropriate for wear-out mechanisms. A constant failure rate does not mean that every item fails at the mean life; it means that age itself does not alter the conditional failure rate.

Weibull model

The Weibull distribution is particularly valuable because its shape parameter can represent several types of ageing:

R(t) = exp[−(t/η)^β],

where β is the shape parameter and η is the scale parameter. The hazard function is

h(t) = (β/η)(t/η)^(β−1).

Interpretation of β is central:

• β < 1 indicates a decreasing failure rate, commonly associated with early-life or commissioning failures.
• β = 1 reduces to the exponential model with constant failure rate.
• β > 1 indicates an increasing failure rate, often associated with wear-out, fatigue, corrosion, or ageing.

The characteristic life η is the time by which approximately 63.2% of the population has failed, because R(η) = e^(−1). It is not the same as MTTF. For a two-parameter Weibull distribution, MTTF = ηΓ(1 + 1/β), where Γ denotes the gamma function.

Normal and lognormal models

The normal distribution may describe dimensions or degradation measurements, but because it permits negative times, it is often unsuitable for time-to-failure unless the mean is much larger than the standard deviation. The lognormal distribution is appropriate when failure time is produced by the multiplicative effect of several factors, such as insulation ageing or crack growth. Its logarithm is normally distributed. It can represent a skewed life distribution, although its hazard behaviour differs from Weibull and must be checked against physical knowledge.

Other distributions

The gamma, generalized gamma, extreme-value, and Poisson-process models may be useful in specialised applications. The choice should not be based solely on the best statistical fit. Physical plausibility, data quality, censoring, the purpose of analysis, and the consequences of model error must also be considered.

The bathtub curve and life-cycle failure

The bathtub curve is a conceptual representation of population failure rate across an asset’s life. It has three regions. The first is infant mortality, where the failure rate decreases. Causes include manufacturing defects, installation errors, design weaknesses, inadequate commissioning, and early contamination. Burn-in, inspection, commissioning tests, and supplier quality improvement may be effective responses.

The second is the useful-life region, where failures occur at an approximately constant rate. Random shocks, accidents, environmental disturbances, and unpredictable component faults dominate. Condition monitoring and good operating controls may reduce consequences, but age-based replacement may not reduce failures if the hazard is genuinely constant.

The third is the wear-out region, where failure rate increases. Fatigue, erosion, corrosion, insulation deterioration, lubrication breakdown, and accumulated damage are typical causes. Replacement, overhaul, refurbishment, redesign, or condition-based intervention may be justified.

The bathtub curve is not a universal law. Different failure modes may have different hazard patterns, and modern electronic systems often do not display a pronounced wear-out phase. A machine-level curve is the combined result of its components, operating conditions, and failure definitions. Treating the curve as proof that all assets should be replaced at a fixed age is a common error.

MTTF, MTBF, MTTR, and availability

Mean time to failure (MTTF) is the expected operating time to failure for a non-repairable item. It is appropriate for a fuse, bearing tested to destruction, or product that is discarded after failure. It is not necessarily the guaranteed life of an item.

Mean time between failures (MTBF) applies to repairable items and measures the mean operating time between successive failures. Under a simple renewal model with statistically stable operating and repair processes, MTBF is the expected uptime interval between repairs. It should not automatically be interpreted as useful life or as a warranty promise.

Mean time to repair (MTTR) is the average active repair time. In practice, restoration time may also include diagnosis, waiting for approval, obtaining parts, travel, setup, testing, and recommissioning. Organisations should specify whether these delays are included. Mean time to restore (MTTR is sometimes used with a broader meaning) may therefore be different from technical repair time.

For a repairable item with stable alternating operating and repair periods, steady-state inherent availability is

Aᵢ = MTBF/(MTBF + MTTR).

If logistics and administrative delays are included, achieved or operational availability should use the corresponding mean uptime and total downtime. For example, an asset with MTBF of 100 hours and MTTR of 5 hours has inherent availability 100/105 = 0.9524, or 95.24%. Reducing MTTR by one hour increases availability more directly than increasing MTBF by a small amount, which illustrates why reliability and maintainability must be managed together.

Reliability for a mission is different from availability. A repairable machine can have high availability while still failing several times during a long production campaign. Conversely, a non-repairable mission component may require high reliability even though repair is impossible during the mission.

Repairable-system measures and event data

For repairable assets, failures and restorations form a recurrent process. The observed failure rate can be estimated as the number of failures divided by total operating time, but this simple estimator assumes comparable exposure, consistent failure definitions, and appropriate treatment of planned downtime and censored observations. Trend analysis is often more informative than one aggregate MTBF. A declining time between failures may indicate deterioration; an apparent improvement may reflect reduced loading, incomplete reporting, or a change in the failure definition.

Reliability-centred maintenance distinguishes functions, functional failures, failure modes, effects, consequences, and technically feasible tasks. Reliability measures should therefore be linked to the consequences of failure rather than pursued as isolated numerical targets. A high MTBF is not automatically desirable if it is achieved by operating below required capacity or by suppressing failure reports.

System reliability: series, parallel, and mixed configurations

A series system fails when any essential component fails. If component failures are statistically independent, system reliability is

R_s(t) = ∏ᵢ Rᵢ(t).

Thus, two components each with reliability 0.98 give series reliability 0.98 × 0.98 = 0.9604. Adding components in series increases functional capability but reduces reliability unless component reliability is improved.

A parallel system succeeds when at least one component operates. For independent components,

R_p(t) = 1 − ∏ᵢ[1 − Rᵢ(t)].

Two components each with reliability 0.90 in active parallel configuration have reliability 1 − (0.10 × 0.10) = 0.99. However, real redundancy may be affected by common-cause failures, shared power, environmental exposure, dependence, imperfect switching, and maintenance errors. These factors can make the actual benefit much smaller than the ideal calculation.

Standby systems contain reserve units that operate only after the primary unit fails. Their reliability depends on switching success, standby exposure, detection time, and whether the standby is cold, warm, or hot. Complex systems can be analysed by reliability block diagrams, fault-tree analysis, Markov models, or simulation. A block diagram represents functional success paths, not necessarily the physical layout of the equipment.

Weibull analysis in practice

Weibull analysis is used to estimate life characteristics from failure and censored data. In a two-parameter model, the linearised relationship is

ln[−ln R(t)] = β ln t − β ln η.

A Weibull probability plot places transformed unreliability against log time. The slope estimates β, while the intercept contributes to η. Modern software generally estimates parameters by maximum likelihood, which is preferable to relying only on graphical fitting, especially with small samples or censored observations.

Right-censored observations occur when a unit has not failed by the end of observation or is withdrawn for another reason. Treating every censored unit as if it failed at the withdrawal time biases the analysis. Suspensions, staggered entry, differing operating hours, and changing conditions must be recorded. Confidence intervals are important because point estimates of β can be highly uncertain when failures are few.

A practical Weibull workflow is:

1. Define the item, mission, operating conditions, and failure criterion.

2. Assemble complete failure and suspension times, including relevant covariates.

3. Separate failure modes where their mechanisms differ materially.

4. Fit plausible distributions and inspect graphical and statistical diagnostics.

5. Interpret β physically rather than mechanically.

6. Estimate reliability at decision-relevant times and quantify uncertainty.

7. Select maintenance, design, or procurement actions and validate them with new data.

For example, β = 0.65 for a group of newly installed pumps suggests decreasing hazard and warrants investigation of installation, alignment, contamination, or commissioning practices. β = 3.2 suggests increasing hazard and supports examination of wear-out mechanisms, inspection intervals, and replacement economics. Neither conclusion alone proves a specific cause; engineering evidence is required.

Applied case: reliability improvement in a packaging line

Background

The following is an illustrative case. A food-packaging plant experienced repeated stoppages on a conveyor drive system. The drive consisted of a motor, gearbox, coupling, and control unit. Production managers reported an availability of approximately 93%, but failure records were inconsistent and did not distinguish mechanical from electrical causes.

Situation and analysis

A six-month data review identified 24 functional failures over 12,000 operating hours. Initial MTBF was therefore approximately 500 operating hours, but this aggregate concealed two different populations. Ten coupling failures occurred mainly during the first 300 hours after replacement, while gearbox failures became more frequent after prolonged service. A Weibull analysis of coupling data produced an estimated β below 1, whereas gearbox data indicated β above 1. The results suggested early-life installation or alignment problems in couplings and wear-out in gearboxes.

The team also measured repair delays. Technical repair time averaged 3 hours, but parts waiting and diagnostic delays increased total downtime to 6 hours. Using MTBF = 500 hours and total restoration time = 6 hours, approximate steady-state availability was 500/(500 + 6) = 98.8% for the simplified failure-repair cycle. The difference between this estimate and the reported 93% showed that planned stops, minor stoppages, data omissions, or a different availability definition were being included in the plant figure.

Approach and solution

The plant introduced a controlled coupling installation procedure, laser alignment verification, torque records, and a 24-hour post-maintenance inspection. For gearboxes, oil analysis and vibration monitoring were introduced, with inspection intervals based on observed degradation rather than a uniform replacement age. A critical spare gearbox was stocked, and the repair kit for couplings was standardised. Failure codes were revised to record mode, cause, operating hours, downtime category, and whether the event was production-critical.

Outcome

In the following operating period, early coupling failures declined, while gearbox interventions increasingly occurred through planned maintenance. The numerical improvement should not be attributed exclusively to the Weibull model: procedural control, better data, and reduced logistics delay also contributed. The central managerial gain was the separation of failure populations and the matching of interventions to their mechanisms.

Learning points

First, a single MTBF can conceal opposing trends. Second, reliability analysis must be connected to physical failure investigation. Third, availability results depend on definitions and data boundaries. Finally, improving maintainability and information quality can create more operational value than pursuing an isolated reliability statistic.

Limitations, risks, and professional judgement

Reliability models are abstractions. Their results can be misleading when data are sparse, failure modes are mixed, operating conditions change, repairs restore items imperfectly, or observations are dependent. Independence assumptions in series and parallel calculations are especially fragile when components share utilities, environments, software, operators, or common causes.

Common mistakes include treating MTBF as guaranteed life, fitting a distribution without checking the failure mechanism, ignoring censored data, combining unlike populations, confusing failure rate with probability, using age replacement despite a constant hazard rate, and reporting excessive precision. Another serious error is excluding inconvenient failures from the database because they are considered operator-caused. Human actions and maintenance quality are part of the operating system and should be analysed rather than hidden.

Reliability information also has ethical and governance implications. Safety-critical failure data must be reported accurately even when disclosure affects production targets or supplier relationships. Models used for safety decisions require traceable assumptions, independent review, version control, and documented uncertainty. Decisions should not rely on algorithmic predictions without engineering validation. Personal data from operator or technician reports should be handled according to organisational privacy requirements.

Digital applications and emerging practice

Computerised maintenance management systems, historians, sensor platforms, and enterprise resource planning systems provide the data infrastructure for reliability analysis. Vibration, temperature, current, pressure, acoustic, and oil-quality measurements can support condition-based maintenance. Statistical learning and machine-learning models may identify abnormal patterns or estimate remaining useful life, but they require representative data, stable sensor quality, explainable decisions, and monitoring for concept drift.

Digital twins and probabilistic risk models increasingly combine design information, operating conditions, inspection results, and maintenance history. Their value lies not in producing a complex model but in improving a decision: whether to inspect, slow down, replace, redesign, or accept risk. Physics-informed models are often more robust than purely data-driven models when failures are rare or historical data are limited.

Best practice is to maintain a common asset hierarchy, standardised failure codes, synchronised time stamps, explicit definitions of uptime and downtime, and a feedback loop from analysis to work execution. Reliability engineering should be integrated with safety, quality, energy, and production objectives rather than treated as a separate reporting activity.

Chapter Summary

Reliability is the probability that an item performs its required function without failure for a specified time and under specified conditions. Reliability, unreliability, density, and hazard functions provide complementary views of failure behaviour. Exponential models assume constant hazard, while Weibull models represent decreasing, constant, or increasing hazard through the shape parameter β. The bathtub curve provides a useful life-cycle concept but is not a universal law.

MTTF applies primarily to non-repairable items; MTBF describes operating intervals of repairable items; MTTR describes restoration effort or time according to the chosen definition. Availability depends on both failure frequency and restoration time. Series systems multiply component reliabilities, whereas parallel systems benefit from redundancy, subject to dependence and common-cause risks. Weibull analysis is most useful when supported by sound failure definitions, censored-data treatment, physical investigation, and uncertainty analysis. Effective maintenance management converts these measures into decisions about design, inspection, replacement, spares, procedures, and risk.

Key Terms

Reliability: Probability that an item performs its required function for a stated time under stated conditions.

Failure: Inability of an item to perform a required function within specified limits.

Fault: A defect or abnormal condition that may or may not have produced functional failure.

Unreliability: Probability that an item fails by a specified time, F(t) = 1 − R(t).

Survival function: Probability that an item survives beyond time t.

Failure density: Distribution of failure times across a population.

Hazard function: Conditional instantaneous failure rate among items surviving to time t.

Cumulative hazard: Integral of the hazard function up to a specified time.

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

MTBF: Mean operating time between successive failures of a repairable item.

MTTR: Mean active repair or restoration time according to the defined measurement boundary.

Availability: Probability or proportion of time that an item is operationally ready.

Exponential distribution: Life distribution with a constant hazard rate.

Weibull distribution: Flexible life distribution whose shape parameter models hazard trends.

Shape parameter β: Weibull parameter indicating decreasing, constant, or increasing hazard.

Scale parameter η: Weibull characteristic life at which about 63.2% have failed.

Series system: System that fails when any essential component fails.

Parallel system: System that succeeds when at least one redundant component operates.

Censoring: Incomplete life observation in which an item has not failed when observation ends.

Common-cause failure: Failure of multiple components due to a shared cause.

Review Questions

Short Answer Questions

1. Distinguish reliability from availability.

2. What does the hazard function represent?

3. State the main interpretations of Weibull shape parameter β.

4. Why is MTBF not a guaranteed operating life?

5. Define a right-censored observation.

Long Answer Questions

1. Explain the relationships among f(t), F(t), R(t), h(t), and H(t).

2. Discuss the three regions of the bathtub curve and appropriate management responses.

3. Compare exponential, Weibull, normal, and lognormal models for failure analysis.

4. Derive and explain reliability equations for independent series and parallel systems.

5. Explain how Weibull analysis should be conducted and interpreted in an industrial setting.

Analytical / Application Questions

1. A repairable machine has MTBF of 240 hours and total restoration time of 8 hours. Calculate its simplified steady-state availability and explain how availability could be improved.

2. Two independent components in series have reliabilities of 0.97 and 0.94 for a mission. Calculate system reliability and identify one design response.

3. A group of replacement bearings shows many early failures, while another group shows failures increasing with age. Propose separate data-analysis and maintenance actions, including the likely Weibull interpretations.

22. Multiple Choice Questions with Answers

Question 1

Which expression correctly defines reliability at time t?

A. P(T ≤ t)

B. P(T > t)

C. f(t)/R(t)

D. 1 + F(t)

Correct Answer: B. P(T > t)

Question 2

A Weibull shape parameter β greater than 1 generally indicates:

A. Decreasing failure rate

B. Constant failure rate

C. Increasing failure rate

D. No failure data

Correct Answer: C. Increasing failure rate

Question 3

An item has MTBF of 100 hours and MTTR of 5 hours. Its simplified inherent availability is closest to:

A. 5.00%

B. 95.24%

C. 97.50%

D. 105.00%

Correct Answer: B. 95.24%

Question 4

For two independent components in series with reliabilities 0.98 and 0.95, system reliability is:

A. 0.030

B. 0.930

C. 0.931

D. 0.998

Correct Answer: C. 0.931

Question 5

For two independent components in active parallel, each with reliability 0.90, system reliability is:

A. 0.81

B. 0.90

C. 0.91

D. 0.99

Correct Answer: D. 0.99

Question 6

Which measure is primarily appropriate for a non-repairable item?

A. MTTF

B. MTBF

C. MTTR

D. Repair frequency

Correct Answer: A. MTTF

Question 7

A constant hazard rate is characteristic of which distribution?

A. Exponential

B. Weibull with β = 2

C. Normal only

D. Lognormal only

Correct Answer: A. Exponential

Question 8

In Weibull analysis, η represents the time by which approximately what proportion has failed?

A. 10.0%

B. 37.0%

C. 63.2%

D. 99.9%

Correct Answer: C. 63.2%

Question 9

Why should censored observations not simply be recorded as failures at the withdrawal time?

A. It makes all failure rates zero

B. It biases the estimated life distribution

C. It converts MTTR into MTBF

D. It eliminates common-cause failures

Correct Answer: B. It biases the estimated life distribution

Question 10

A machine shows early coupling failures and late gearbox failures. The most appropriate response is to:

A. Apply one uniform replacement interval to all components

B. Ignore failure modes and use aggregate MTBF

C. Analyse the populations separately and match actions to mechanisms

D. Assume all failures are random

Correct Answer: C. Analyse the populations separately and match actions to mechanisms