System Reliability Engineering, Failure Rate Physics & Accelerated Life Testing

Reliability engineering applies probability theory, stochastic processes, and material physics to analyze the time-to-failure of components, subsystems, and integrated industrial architectures.

Let the operating lifetime of an engineering component or system be modeled as a continuous, non-negative random variable $T \ge 0$ defined on a probability space $(\Omega, \mathcal{F}, P)$. The probability distribution of $T$ is characterized by several interrelated mathematical functions:

1. Cumulative Distribution Function (Unreliability Function)

The Unreliability Function $F(t)$ represents the probability that the component fails at or before operating time $t$:

where $f(t) = \frac{d F(t)}{dt}$ is the failure probability density function (PDF).

2. Survival Function (Reliability Function)

The Reliability Function $R(t)$ represents the probability that the component survives beyond operating time $t$ without failure:

Mathematical boundary conditions require $R(0) = 1.0$ (perfect reliability at inception) and $\lim_{t \to \infty} R(t) = 0$ (all physical systems eventually fail).

3. Hazard Rate Function (Instantaneous Failure Rate)

The Hazard Rate Function $h(t)$ (also denoted $\lambda(t)$) represents the conditional failure rate per unit time, given that the component has survived up to time $t$:

Applying conditional probability:

Dividing by $\Delta t$ and taking the limit:

4. Cumulative Hazard Function

Integrating the hazard rate from $0$ to $t$ yields the Cumulative Hazard Function $H(t)$:

Exponentiating both sides yields the foundational relationship connecting reliability directly to cumulative hazard:

The failure PDF can then be expressed exclusively in terms of hazard:

These identities demonstrate that specifying any one of the four functions ($f(t), F(t), R(t), h(t)$) completely and uniquely determines the remaining three.

Mean Time To Failure (MTTF)

For non-repairable systems, the expected operating lifetime is the Mean Time To Failure (MTTF):

Applying integration by parts with $u = t$ and $dv = f(t)dt \implies v = -R(t)$:

Assuming $\lim_{t \to \infty} t R(t) = 0$ (which holds for all physically realistic lifetime distributions with finite mean):

The expected lifetime equals the total area under the reliability curve!

Mean Time Between Failures (MTBF)

For repairable systems (systems restored to operational status upon failure via maintenance or component replacement), the operational metric is the Mean Time Between Failures (MTBF):

where $\text{MTTR}$ is the Mean Time To Repair. If repair time is negligible compared to operating duration ($\text{MTTR} \ll \text{MTTF}$), $\text{MTBF} \approx \text{MTTF}$.

Mean Residual Life (MRL)

The expected additional operating lifespan remaining for a component that has already survived up to time $t$ is the Mean Residual Life $MRL(t)$:

If the system experiences physical wear-out (increasing hazard rate), $MRL(t)$ decreases monotonically with operating age.

1. The Exponential Lifetime Distribution

The Exponential distribution models systems subject purely to random, external environmental shocks with zero aging:

• Failure PDF: $f(t) = \lambda e^{-\lambda t}, \quad t \ge 0, \; \lambda > 0$
• Reliability Function: $R(t) = e^{-\lambda t}$
• Hazard Rate Function: $h(t) = \frac{f(t)}{R(t)} = \frac{\lambda e^{-\lambda t}}{e^{-\lambda t}} = \lambda \quad (\text{constant!})$
• $\text{MTTF} = \int_0^\infty e^{-\lambda t}\,dt = \frac{1}{\lambda}$

The defining mathematical property of the exponential distribution is its Memoryless Property:

A component that has operated for $10,000\text{ hours}$ without failing is probabilistically as good as a brand-new component fresh off the assembly line. While applicable to electronic semiconductor components, the exponential model is invalid for mechanical components that experience wear, fatigue, or corrosion.

2. The Two-Parameter Weibull Distribution

Invented by Waloddi Weibull in 1951, the Weibull distribution is the most versatile parametric distribution in reliability engineering, capable of modeling increasing, decreasing, or constant failure rates:

• Reliability Function:

• Cumulative Hazard: $H(t) = (t / \eta)^\beta$
• Hazard Rate Function:

• Failure PDF:

where:

• $\beta > 0$: The shape parameter (Weibull slope), which governs the physics of failure:
• $\beta < 1$: Decreasing Failure Rate (DFR). Models "infant mortality" or burn-in defects caused by manufacturing flaws, poor soldering, or material contaminants.
• $\beta = 1$: Constant Failure Rate (CFR). The Weibull collapses exactly to the exponential distribution with $\lambda = 1/\eta$.
• $\beta > 1$: Increasing Failure Rate (IFR). Models physical wear-out, mechanical fatigue, erosion, and chemical degradation.
• $\beta \approx 3.44$: The Weibull distribution becomes nearly symmetric, closely approximating a Gaussian distribution.
• $\eta > 0$: The scale parameter (Characteristic Life). Notice that at $t = \eta$:

The characteristic life $\eta$ is the exact operating time by which $63.2\%$ of the population will have failed, regardless of the value of shape parameter $\beta$!

The Mean Time To Failure of a Weibull component is evaluated via the Gamma function:

3. The Bathtub Curve

The synthesis of these mechanisms over the full product lifecycle creates the celebrated Bathtub Curve of failure rates:

1. Infant Mortality Phase (Early Life): $h(t)$ decreases rapidly ($\beta < 1$) as latent manufacturing defects are weeded out during factory burn-in screening.

2. Useful Life Phase (Intrinsic Life): $h(t)$ remains approximately flat and constant ($\beta \approx 1$), dominated by random external operating shocks. The exponential model governs this phase.

3. Wear-Out Phase (End of Life): $h(t)$ accelerates upward ($\beta > 1$) due to cyclic fatigue, oxidation, and mechanical friction. Preventative maintenance replaces components before this phase begins.

Complex industrial systems are networks of interconnected components. Let a system consist of $k$ components with individual reliabilities $R_1(t), R_2(t), \dots, R_k(t)$, operating independently.

1. Series System Architecture (Weakest-Link Principle)

In a series system, the failure of any single component causes catastrophic failure of the entire system (e.g., a holiday string of lights wired in series, or a multi-stage rocket powertrain):

Taking logarithms:

Differentiating with respect to $t$ yields the additive property of hazard rates:

If all $k$ components have constant failure rates $\lambda_i$, the series system has an exponential lifetime with failure rate $\lambda_{\text{series}} = \sum_{i=1}^k \lambda_i$, and:

Notice that $R_{\text{series}}(t) \le \min_i(R_i(t))$: a series system is always less reliable than its least reliable component!

2. Parallel Redundant Architecture

In a parallel system, the system operates as long as at least one component survives. The system fails only if all $k$ components fail simultaneously (e.g., dual avionics flight computers, or twin aircraft jet engines):

For $k$ identical components with exponential failure rate $\lambda$:

The MTTF of an identical $k$-component parallel system is:

Parallel redundancy provides massive reliability gains, but with diminishing returns as $k$ expands.

3. $k$-out-of-$n$ Voting Systems

A $k$-out-of-$n$ system contains $n$ identical, independent components and functions if and only if at least $k$ components remain operational (e.g., a quad-engine commercial airliner that can fly safely on at least 2 engines, or a triple-modular redundant (TMR, 2-out-of-3) nuclear reactor shutdown circuit).

Using the Binomial probability distribution:

For a 2-out-of-3 system:

Reliability life testing rarely observes every unit until failure. Running tests until 100% of units fail takes years and wastes resources. Instead, tests are terminated under censoring protocols:

• Type I (Time) Censoring: The life test is stopped at a pre-determined fixed calendar time $\tau$. The number of observed failures $r$ is a random variable.
• Type II (Failure) Censoring: The life test is stopped as soon as a pre-determined number of failures $r$ has occurred ($r < n$). The termination time $T_{(r)}$ is a random variable.
• Random Right Censoring: Units enter testing at different dates or drop out due to competing un-related failures.

The Likelihood Construction

Let $n$ components be placed on test. For component $i$, we observe time $t_i$ and failure indicator $\delta_i \in \{0, 1\}$:

The contribution of an uncensored failure to the likelihood is the PDF $f(t_i)$. The contribution of a censored survival is the probability of surviving past $t_i$, which is $R(t_i)$. The joint likelihood function is:

The log-likelihood is:

Exponential Distribution with Type II Censoring

For an exponential distribution with hazard rate $\lambda$, test stopped at the $r$-th failure among $n$ units ($t_{(1)} \le t_{(2)} \le \dots \le t_{(r)}$): The total time on test (TTT) accumulated by all $n$ units is:

The log-likelihood is:

Differentiating and setting to zero yields the Maximum Likelihood Estimator:

Benjamin Epstein and Milton Sobel (1953) proved that $2 \lambda T_{\text{total}} \sim \chi^2_{2r}$. An exact two-sided $(1 - \alpha)100\%$ confidence interval for the true MTTF is:

In modern electronics and photonics, high-reliability components have expected lifetimes exceeding $500,000\text{ hours}$ (over 50 years). Testing such devices under nominal operating conditions ($25^\circ\text{C}, 1.0\text{ V}$) will observe zero failures. Accelerated Life Testing (ALT) exposes units to severe elevated stresses (elevated temperature, voltage, mechanical vibration, humidity) to accelerate the underlying chemical and physical degradation mechanisms, forcing failures to occur within days or weeks.

The fundamental premise of ALT is that the acceleration stress increases the rate of the failure reaction without altering the underlying physical failure mode (the shape parameter $\beta$ remains invariant across stress levels).

1. Thermal Acceleration: The Arrhenius Model

Svante Arrhenius (1889) formulated chemical reaction kinetics as a function of thermal activation energy. In reliability physics, the reaction rate $K(T)$ at absolute temperature $T$ (in Kelvin, $K = ^\circ\text{C} + 273.15$) is:

where:

• $E_a$: Activation energy of the failure mechanism in electron-volts ($\text{eV}$), typically $0.3\text{ eV}$ to $1.2\text{ eV}$ (e.g., $E_a \approx 0.7\text{ eV}$ for silicon oxide dielectric breakdown).
• $k_B$: Boltzmann's constant $= 8.61733 \times 10^{-5}\text{ eV/K}$.
• $T$: Absolute temperature in Kelvin.

Because operating lifetime $t$ is inversely proportional to reaction rate $K$, the Thermal Acceleration Factor ($AF_{\text{thermal}}$) between operational temperature $T_{\text{use}}$ and elevated stress temperature $T_{\text{stress}}$ is:

2. Voltage Acceleration: The Inverse Power Law

For dielectric breakdown in ceramic capacitors and gate oxide wear-out:

where $n$ is a voltage acceleration exponent (typically $n \in [3, 7]$).

3. Temperature-Humidity Acceleration: Peck's Model

For epoxy package corrosion and ionic contamination under moisture:

where $RH$ is relative humidity and $m \approx 2.7$ to $3.0$.

Once the acceleration factor $AF$ is established, the characteristic life at nominal use conditions is extrapolated: $\eta_{\text{use}} = AF \cdot \eta_{\text{stress}}$, enabling rigorous lifetime guarantees from short-duration testing.

In complex mission-critical systems (e.g., nuclear power plant safety injection systems, space exploration launch vehicles, and chemical plant flare headers), reliability cannot be represented as simple series-parallel combinations. Engineers utilize Fault Tree Analysis (FTA) and Reliability Block Diagrams (RBD) to model catastrophic system failure.

Fault Tree Methodology and Minimal Cut Sets

A Fault Tree is a top-down deductive graphical failure model connecting an undesired catastrophic Top Event (e.g., "Complete Loss of Flight Control") to constituent basic hardware, software, and human failures via Boolean logic gates (AND, OR, Voting gates).

Let $X_i \in \{0, 1\}$ denote the binary state of basic event $i$:

The state of the Top Event is governed by the system structure function $\psi(\mathbf{X}) \in \{0, 1\}$. A Cut Set is a set of basic events whose joint occurrence forces $\psi(\mathbf{X}) = 1$ (causes top event failure). A Minimal Cut Set (MCS) is a cut set that ceases to be a cut set if any single basic event is removed. The minimal cut sets represent the smallest combinations of primary failures sufficient to crash the entire system.

By applying Boolean algebra (De Morgan’s laws and absorption laws), any complex fault tree can be reduced to an equivalent sum of minimal cut sets $C_1, C_2, \dots, C_m$:

The exact Top Event probability $Q_{\text{sys}} = P(\psi(\mathbf{X}) = 1)$ is evaluated via the Poincaré Inclusion-Exclusion formula or upper-bounded by the union bound:

where $q_i = P(X_i = 1)$ is the failure probability of basic event $i$.

Component Reliability Importance Measures

To determine which individual components contribute most critically to overall system vulnerability, reliability engineers compute quantitative importance indices:

1. Birnbaum Structural Importance ($I_B(i)$):

The probability that component $i$ is critical to system failure (i.e., the system operates when $i$ works, but crashes the instant $i$ fails):

2. Fussell-Vesely Importance ($I_{FV}(i)$):

The fractional contribution of component $i$ to total system unreliability, measuring the probability that component $i$ caused the system failure given that the system has failed:

Components exhibiting high Birnbaum or Fussell-Vesely scores are targeted for redundancy, improved maintenance scheduling, or aerospace-grade component upgrades.

Beyond empirical statistical curve fitting, modern reliability engineering integrates material physics and solid mechanics under the Physics-of-Failure (PoF) paradigm.

Thermomechanical Fatigue: The Coffin-Manson Model

In electronic assemblies and aerospace turbine blades, cyclic heating and cooling causes mismatched thermal expansion between adjacent materials (e.g., silicon die, copper leadframe, and solder alloy). This cyclic thermal strain produces low-cycle plastic deformation and fatigue cracking.

L. F. Coffin and S. S. Manson (1953) formulated the relationship between the cyclic plastic strain range $\Delta \varepsilon_p$ and the number of cycles to failure $N_f$:

where $\varepsilon_f'$ is the fatigue ductility coefficient and $c \approx -0.5$ to $-0.7$ is the fatigue ductility exponent.

In electronics thermal shock testing, the Modified Coffin-Manson Acceleration Factor between operational temperature cycle $\Delta T_{\text{use}}$ and test cycle $\Delta T_{\text{stress}}$ is:

where:

• $\Delta T = T_{\text{max}} - T_{\text{min}}$: The thermal cycle amplitude.
• $q$: The Coffin-Manson temperature exponent (typically $q \approx 1.9$ to $2.5$ for lead-free SAC305 solder).
• $f$: Cycling frequency, accounting for solder creep stress relaxation.
• $E_a$: Activation energy for thermal creep diffusion ($\sim 0.12\text{ eV}$).

The Coffin-Manson equation enables automotive and aerospace engineers to simulate 15 years of daily ambient automotive thermal cycling ($25^\circ\text{C}$ to $125^\circ\text{C}$) in only 3 weeks of rapid environmental chamber cycling ($-40^\circ\text{C}$ to $+150^\circ\text{C}$).

In complex industrial machinery, operating lifetimes are rarely governed by calendar time alone. Component degradation is heavily modulated by dynamic operational covariates: mechanical load, rotational speed, oil contamination levels, vibration amplitude, and thermal cycling.

The Cox Proportional Hazards Model for Reliability

In 1972, Sir David Cox formulated the Proportional Hazards Model, which separates baseline aging from operational stress covariates:

where:

• $h_0(t)$: The baseline hazard function, representing the intrinsic failure rate of an asset operating under standard baseline conditions ($\mathbf{z} = \mathbf{0}$).
• $\exp(\boldsymbol{\beta}^T \mathbf{z})$: The relative risk multiplier, which scales the failure hazard proportionally based on environmental stress.

If the baseline hazard follows a Weibull distribution ($h_0(t) = \frac{\beta_0}{\eta_0} (t/\eta_0)^{\beta_0 - 1}$), the model becomes the Weibull Proportional Hazards Model:

This formulation enables predictive maintenance engineers to compute the dynamic Remaining Useful Life (RUL) of an operational asset conditioned on real-time SCADA telemetry sensor readings.

Competing Risks in Multimodal Failure Systems

An industrial machine often fails due to one of several mutually exclusive physical failure mechanisms (e.g., a jet turbine blade may fail due to high-cycle fatigue cracking, thermal oxidation, or foreign object damage impact). These distinct mechanisms compete to cause the failure of the asset.

Let $J \in \{1, 2, \dots, K\}$ denote the failure mode. The Cause-Specific Hazard Function for failure mode $j$ is:

The overall all-cause hazard rate of the system is the sum of cause-specific hazards:

The overall system survival function is:

The probability that failure occurs specifically due to mechanism $j$ on or before time $t$ is the Cumulative Incidence Function (CIF):

Analyzing cause-specific hazards prevents engineers from treating competing failure modes as independent events, ensuring that design modifications targeting one failure mode properly anticipate the emergence of latent secondary modes.

In ultra-high reliability engineering—such as space mission satellites, commercial aircraft avionics, and deep-well subsea blow-out preventers—qualification tests often terminate with zero failures observed ($r = 0$). Under classical Maximum Likelihood Estimation, if $r = 0$, the point estimate for failure rate is $\widehat{\lambda} = 0 / T_{\text{total}} = 0$, implying infinite MTTF—a physical absurdity that cannot guide engineering decisions.

To extract rigorous statistical inference from zero-failure testing, reliability engineers utilize Bayesian Reliability Methods, integrating prior engineering physics and historical component data with test observations.

Conjugate Gamma-Exponential Model

Let the component lifetime follow an exponential distribution with failure rate $\lambda \sim \text{Exp}(\lambda)$. The prior knowledge regarding $\lambda$ is modeled using a conjugate Gamma Distribution $\text{Gamma}(\alpha_0, \beta_0)$:

where $\alpha_0$ represents the prior pseudo-failure count and $\beta_0$ represents the prior pseudo-operating hours.

Suppose an empirical life test observes $r$ failures accumulated over total time on test $T_{\text{total}}$. By Bayes' theorem, the posterior distribution is also Gamma:

Thus, the posterior distribution is:

Posterior Point and Interval Estimators

The posterior expected failure rate and posterior MTTF are:

When zero failures occur ($r = 0$) under an uninformative prior ($\alpha_0 = 1, \beta_0 = 0$, corresponding to a uniform prior on $\ln \lambda$):

Furthermore, a one-sided $(1 - \alpha)100\%$ Bayesian lower credibility limit on reliability $R(t) = e^{-\lambda t}$ is derived directly from the cumulative chi-square distribution:

Bayesian reliability analysis converts zero-failure test runs into valid contractual qualification certificates, preventing expensive, pointless test overruns.