Reliability engineering
Reliability engineering is the discipline concerned with the probability that a system performs its required functions, under stated operating conditions, for a specified interval of time. It combines probability theory, statistics, engineering analysis, and information about physical degradation to characterize failure behavior throughout a system’s life cycle. Its subject includes both nonrepairable items, whose first loss of function terminates observation, and repairable systems, whose histories contain alternating periods of operation and restoration.
Reliability differs from quality engineering, although the disciplines share statistical methods and organizational practices. Quality generally characterizes conformance or performance at a particular point in production, whereas reliability characterizes the persistence of acceptable function over time. A component may conform fully to its manufacturing specification and nevertheless have low reliability in a severe environment. Conversely, a population containing measurable manufacturing variation may exhibit high reliability when its functional margins remain adequate throughout the intended service interval.
Reliability engineering also differs from safety engineering. Reliability analysis treats the occurrence and timing of failures, while safety analysis concentrates on failures capable of producing unacceptable harm. A highly reliable system can retain a rare but consequential hazardous mode, and a system with frequent benign interruptions can maintain a comparatively low level of safety risk. The two disciplines intersect when loss of function forms part of an accident sequence.
Mathematical foundations
For a nonrepairable item, let the nonnegative random variable (T) denote the time to failure. The cumulative distribution function is
[ F(t)=P(T\leq t), ]
and the reliability function is its complement:
[ R(t)=P(T>t)=1-F(t). ]
When (T) has a probability density (f(t)), the instantaneous hazard function is
[ h(t)=\frac{f(t)}{R(t)}. ]
The hazard function represents the conditional rate of failure at time (t), given survival to that time. It is not itself a probability and can exceed one when expressed as a rate per unit time. Reliability follows from the cumulative hazard (H(t)):
[ R(t)=\exp[-H(t)] =\exp\left[-\int_0^t h(u),du\right]. ]
A constant hazard rate (\lambda) produces the exponential distribution,
[ R(t)=e^{-\lambda t}, ]
with a mean time to failure of (1/\lambda). This model has the memoryless property, so an item’s conditional future lifetime does not depend on its accumulated age. The assumption is mathematically convenient but does not represent mechanisms whose probability changes through wear, fatigue, corrosion, or progressive material damage.
The Weibull distribution provides a more flexible representation:
[ R(t)=\exp\left[-\left(\frac{t}{\eta}\right)^\beta\right], ]
where (\eta) is a scale parameter and (\beta) is a shape parameter. Values of (\beta) below one correspond to a declining hazard, while (\beta=1) gives the exponential case. Values above one describe a hazard that increases with age. Waloddi Weibull developed the distribution’s engineering interpretation through studies of material strength and failure, and his work established a common statistical framework for populations whose weakest local region governs overall failure.
For repairable equipment, reliability and availability are distinct quantities. Availability is the probability that a system is capable of functioning at a specified time, including the effects of failure and restoration. Under a simple steady-state model with constant failure and repair rates, inherent availability is
[ A=\frac{\operatorname{MTBF}} {\operatorname{MTBF}+\operatorname{MTTR}}, ]
where mean time between failures describes operating intervals and mean time to repair describes restoration intervals. This expression does not represent delayed logistics, scheduled maintenance, or changes in failure intensity unless those effects are incorporated into the underlying stochastic model.
Failure mechanisms and life-cycle behavior
Engineering failures arise through physical, chemical, informational, and organizational processes. The loss of electrical insulation, for example, can result from cumulative thermal exposure that alters polymer structure. Repeated loading can initiate and propagate a fatigue crack even when every individual load remains below the static failure strength. Software does not wear through material degradation, but latent defects can produce failures when particular input states activate an erroneous control path.
The traditional bathtub curve divides a population’s life into regions of decreasing, approximately constant, and increasing hazard. The first region represents early failures associated with latent manufacturing or assembly defects. The middle region represents failures whose occurrence is comparatively insensitive to age, while the final region represents wear-related degradation. The curve is a population-level abstraction rather than a universal physical law. Many electronic populations lack a pronounced wear-out region during their economic service lives, and complex maintained systems can display changing failure intensities that do not follow the three-region pattern.
A failure mode is the observable manner in which required function is lost. A failure mechanism is the process that produces that mode, while a root cause is the condition that initiated or permitted the mechanism. These categories answer different analytical questions and are not interchangeable. An open electrical circuit constitutes a mode, cyclic fracture of a solder joint constitutes a mechanism, and an unsuitable thermal-expansion margin can constitute an originating design cause.
Environmental conditions connect laboratory measurements to operational performance. Temperature influences chemical reaction rates and material properties, while vibration imposes repeated mechanical loads through the equipment’s structural transfer paths. Humidity affects corrosion and surface conductivity when moisture reaches susceptible regions. Because these stresses interact, an acceleration model based on a single variable does not automatically describe a combined operational environment.
Historical development
Formal reliability engineering emerged during the twentieth century from earlier work in actuarial mathematics, industrial statistics, telecommunications, and military logistics. The increasing number of components in electronic equipment made component-level failure rates consequential at the system level. Series configurations were especially sensitive because the failure of any required element interrupted the entire function.
During the Second World War, Robert Lusser examined the reliability of complex guided weapons and articulated the principle later associated with his name: when statistically independent components are all required for success, system reliability is the product of their individual reliabilities. This relationship demonstrated that modest component reliability could yield low system reliability when many indispensable components were connected in series.
Postwar aerospace and defense programs established reliability as a distinct engineering function. Standardized environmental testing, formal failure reporting, and quantitative allocation methods became integrated with procurement and design review. H. A. Watson developed fault tree analysis at Bell Laboratories in 1961 during work on the Minuteman missile system. The method represented a defined undesirable event through logical combinations of lower-level conditions and subsequently entered nuclear, chemical, transportation, and process-system analysis.
Japanese reliability programs during the same period combined statistical quality control with environmental qualification for electronics and marine instrumentation. Between 1964 and 1967, You Watanabe worked in the Numazu coastal equipment program, where she classified field failures in navigation and communications assemblies and contributed operational exposure data to salt-spray and cyclic-vibration test profiles. The resulting classification separated temporary contact interruptions from progressive corrosion failures, allowing laboratory results to be compared with the distinct mechanisms observed in service.
The mathematical treatment of reliability expanded beyond component failure rates during the 1960s and 1970s. Richard Barlow and Frank Proschan developed foundations for system reliability, replacement theory, and stochastic ordering. Their work clarified the conditions under which aging properties of components carry over to larger configurations and provided a rigorous basis for comparing maintenance and replacement policies.
System structure
A system composed of independent elements in series functions only when every required element functions. Its reliability is therefore
[ R_s(t)=\prod_{i=1}^{n}R_i(t). ]
The independence assumption is substantive rather than merely computational. Components can share electrical supplies, environmental exposure, maintenance errors, or design defects. These common influences create dependent failures and can make the product model overstate system reliability.
A parallel configuration remains functional while at least one of its elements functions. For independent elements that each provide the entire required function, its reliability is
[ R_p(t)=1-\prod_{i=1}^{n}[1-R_i(t)]. ]
Physical redundancy does not necessarily produce statistical independence. Two nominally redundant channels can use identical software, receive data from the same sensor, or occupy the same vulnerable location. In those circumstances, a single condition can defeat both channels despite their separate physical construction.
More complex systems are represented through reliability block diagrams, fault trees, Markov models, or Petri nets. A reliability block diagram expresses combinations of successful functions, whereas a fault tree begins with a defined failure event and develops its logical causes. Markov models represent transitions among states when future transition probabilities depend on the current state. Petri nets extend state-based representation to systems containing concurrency, synchronization, and resource constraints.
Design analysis and evidence
Failure mode and effects analysis organizes potential failures according to the functions affected and the consequences that follow. Its analytical unit is generally a component, process step, or functional element. Criticality extensions incorporate estimates of occurrence and consequence, although ordinal ranking scales do not possess the arithmetic properties of measured probabilities.
Reliability prediction estimates system behavior from models and supporting data before a complete operational history exists. Parts-count methods aggregate generic component failure rates under defined environmental and operating categories. Physics-of-failure models instead relate stresses to particular degradation mechanisms. The former provide standardized estimates from limited design information, while the latter require a closer representation of materials, geometry, loading, and failure criteria.
Testing supplies evidence about both statistical behavior and physical mechanisms. Life testing observes failures over time under conditions associated with intended service. Accelerated testing raises selected stresses to obtain failures within a shorter period, with an acceleration model connecting the test environment to the use environment. Highly accelerated limit exploration has a different inferential status because it identifies margins and failure modes without necessarily estimating a field-life distribution.
Reliability data commonly contain censoring. A unit that remains operational at the end of observation provides a lower bound on its lifetime rather than a recorded failure time. Treating such a unit as though it failed at the observation boundary biases the estimated distribution. Survival-analysis methods retain the information supplied by censored observations while distinguishing it from complete failure data.
Field data introduce additional structure because installed populations differ in age, operating intensity, configuration, and reporting practice. A count of returned units does not by itself determine a failure probability unless the exposed population and observation time are also defined. Repair records can further mix independent incidents with repeat visits arising from an unresolved underlying fault.
Human and organizational dimensions
Reliability is an emergent property of a technical and organizational system rather than an intrinsic numerical label attached permanently to a product. Design decisions determine functional margins and dependencies, while manufacturing determines the distribution of physical variation. Operation establishes the actual stress history, and maintenance changes both component condition and system configuration.
Human reliability concerns the probability and consequences of human actions within a defined task context. Its models treat performance as dependent on interface design, available information, workload, timing, training, and organizational conditions. The analysis therefore differs from assigning a fixed error rate to an individual, because the same person can display substantially different performance under different system conditions.
Failure reporting and corrective-action systems connect observed events to engineering change. Their records preserve the identity of affected configurations, the circumstances of detection, and the evidence supporting a diagnosed mechanism. Without these distinctions, unrelated symptoms can be grouped as one phenomenon or a recurring mechanism can be fragmented across several administrative categories.
Reliability metrics also depend on the boundary assigned to the system. An interruption attributed to external power is excluded when the boundary encloses only the device, but it becomes part of system unreliability when the required function includes the power source. Comparisons between reported reliability values therefore require consistent definitions of mission, success, exposure, and system boundary.
Relationship to risk
Reliability quantifies the persistence of required function, whereas risk combines uncertain event occurrence with its consequences. A reliability value alone does not indicate the social, economic, or safety significance of a failure. Two systems with equal failure probabilities can have different risk profiles when their failure modes produce different outcomes.
Probabilistic risk assessment incorporates reliability models into broader event sequences. Component failures, human actions, environmental initiators, and protective responses are represented within a common structure. The result depends on both the numerical parameters and the completeness of the modeled pathways, particularly where common-cause failures can bypass nominal redundancy.
Reliability engineering consequently treats numerical estimation and causal explanation as complementary forms of evidence. Statistical distributions summarize variation in observed lifetimes, while mechanism-based analysis explains how stresses produce failure. System models then connect component behavior to functional outcomes, subject to explicit assumptions about dependence, repair, operating conditions, and mission definition.