Weibull Distribution

The shape parameter β (beta) is the defining characteristic of the Weibull distribution, directly mapping to the bathtub curve of asset reliability:

• β < 1: Models infant mortality or early-life failures. The hazard rate decreases over time as weak components are eliminated. This is typical of new transformers with manufacturing defects.
• β = 1: Reduces to the exponential distribution, modeling random failures during the useful life period. Failures occur at a constant rate due to external events like lightning strikes or switching surges.
• β > 1: Models wear-out failures. The hazard rate increases over time as insulation degrades and mechanical components fatigue. A β of 3-4 is common for transformer paper insulation aging.
• β > 10: Indicates rapid, deterministic wear-out, approaching a normal distribution.

The scale parameter η (eta) defines the characteristic life of the asset population. It is the time at which 63.2% of the population will have failed, regardless of the shape parameter value.

• η is expressed in the same units as the time-to-failure data, typically hours or years for transformers.
• For a given β, a larger η indicates a more reliable population with longer expected life.
• In transformer fleet analysis, η helps asset managers benchmark different sub-populations, such as comparing the characteristic life of transformers from different manufacturers or those operating under different loading profiles.
• The relationship between η and the Mean Time To Failure (MTTF) is: MTTF = η × Γ(1 + 1/β), where Γ is the gamma function.

The Probability Density Function (PDF) describes the relative likelihood of failure at any given time t. Its shape is dictated entirely by β:

• For β < 1: The PDF is monotonically decreasing, with the highest probability of failure at time zero, decreasing thereafter.
• For β = 1: The PDF is a simple exponential decay curve.
• For β > 1: The PDF is unimodal and right-skewed, peaking at the mode of the distribution. As β increases, the distribution becomes more symmetric and concentrated around η.
• In Dissolved Gas Analysis (DGA)-driven predictive maintenance, the PDF helps quantify the probability that a transformer with a specific gas trend will fail within the next inspection interval, enabling risk-based maintenance scheduling.

The reliability function R(t) gives the probability that a transformer survives beyond time t. It is the complement of the cumulative distribution function: R(t) = e^-(t/η)^β.

• At t = η, R(t) = e^-1 ≈ 36.8%, confirming that 63.2% have failed by the characteristic life.
• This function is critical for Remaining Useful Life (RUL) estimation. Given a transformer has survived to its current age, the conditional reliability function calculates the probability of surviving an additional mission time.
• Asset managers use R(t) curves to determine optimal replacement intervals, balancing the cost of unplanned failure against the cost of proactive replacement.
• The Bx life (e.g., B10 life) is derived directly from R(t), indicating the time by which x% of the population is expected to fail.

The hazard function h(t) represents the instantaneous failure rate at time t, given survival to that time. Its formula is: h(t) = (β/η) × (t/η)^(β-1).

• This is the most diagnostically useful function for Condition-Based Maintenance (CBM). A rising h(t) signals accelerating degradation.
• For transformer windings, the hazard rate often follows a power-law relationship with Hot-Spot Temperature, which can be integrated into a Physics-Informed Neural Network (PINN) to constrain failure predictions with Arrhenius-based thermal aging laws.
• The cumulative hazard function H(t) = (t/η)^β is used in Weibull probability plotting to visually assess goodness-of-fit on Weibull probability paper.
• A flat hazard function (β=1) implies that preventive replacement offers no reliability improvement over run-to-failure.

Accurate estimation of β and η from field data is essential for valid predictions. Two primary methods are used:

• Maximum Likelihood Estimation (MLE): Solves for parameters that maximize the probability of observing the actual failure and suspension data. MLE handles right-censored data (transformers still operating) naturally, making it ideal for fleets with incomplete failure histories.
• Rank Regression on X (RRX): Fits a linear model to the transformed data on Weibull probability paper using median ranks (often approximated by Benard's formula: (i-0.3)/(n+0.4)). RRX is robust for small sample sizes.
• Bayesian Weibull analysis incorporates prior engineering knowledge—such as the expected β for paper insulation aging—to refine estimates when failure data is sparse.
• Confidence bounds on parameters are critical; a wide confidence interval on β that spans 1.0 indicates insufficient data to distinguish random from wear-out failure modes.

A prognostic metric estimating the operational time left before a transformer asset degrades to a predefined failure threshold. Weibull distribution provides the foundational hazard rate function that drives RUL calculations by modeling how failure probability evolves over time. RUL estimation combines the shape parameter (β) from Weibull analysis with real-time condition monitoring data such as dissolved gas trends and hot-spot temperatures to project end-of-life with confidence intervals.

A maintenance strategy that uses real-time sensor data and diagnostic indicators to schedule repairs only when evidence of decreasing equipment performance is detected. Weibull analysis underpins CBM by establishing the conditional probability of failure given current asset age and condition. This shifts maintenance from fixed calendar intervals to risk-based scheduling, where the Weibull hazard function determines the optimal intervention point before the failure probability accelerates unacceptably.

A composite numerical score calculated by weighting multiple diagnostic test results and operational history to provide a simplified overall condition ranking for a transformer fleet. Weibull parameters are often used to calibrate the weighting factors within health index models by correlating historical failure data with diagnostic measurements. A declining health index mapped against the Weibull cumulative distribution function enables asset managers to prioritize replacements based on statistical failure risk rather than heuristic scoring alone.

The supervised machine learning task of categorizing transformer fault types based on labeled patterns in dissolved gas and electrical test data. Weibull analysis complements classification by providing failure mode-specific shape parameters:

• β ≈ 1: Random failures (lightning strikes, external faults)
• β > 2: Wear-out failures (insulation aging, contact erosion)
• β < 1: Manufacturing defects (early-life dielectric weaknesses) This stratification enables separate reliability models for each failure mechanism.

The application of statistical or deep learning models like LSTM and Temporal Fusion Transformer to predict future gas levels or temperature trajectories. Weibull distribution serves as the probabilistic output layer for many forecasting models, converting predicted degradation trajectories into time-to-failure probability distributions. This hybrid approach combines the pattern recognition power of neural networks with the statistical rigor of reliability engineering for defensible maintenance recommendations.

A machine learning technique that combines multiple predictive models to improve fault classification accuracy. In transformer prognostics, ensemble methods often incorporate Weibull-based survival probabilities as input features alongside dissolved gas ratios and thermal data. The Weibull hazard function acts as a domain-knowledge feature that constrains purely data-driven models to respect the bathtub curve behavior observed in real transformer populations, reducing false alarms.