Degradation Path Analysis

Description

Model each unit's repeated measurements of decline with a random-coefficient curve, define failure as crossing a threshold, and derive the distribution of crossing times.

Many failures are preceded by measurable decline, and that decline carries far more information than the few failures seen in a short study. Degradation path analysis models each unit's measurements over time as a smooth path with a few parameters, some varying randomly between units, plus measurement error; scales are chosen so paths are simple, often linear in logs. A soft failure is defined as the path reaching a critical level. The time-to-failure distribution is the probability that a path has crossed the level by each time, computed from the fitted model by simulation, with bootstrap intervals. A quick approximation fits each unit's path separately, extrapolates to the threshold to get pseudo failure times, and analyses those as ordinary failure data. It works only when the measured quantity is closely tied to the outcome of interest.

Attributes

Loss type
Unmigrated
Stage
Post-launch
Effort
Weeks

Steps

  1. Choose a measurable quantity closely related to the outcome (for example weekly active minutes or a skill score) and a critical level that defines failure.
  2. Measure each unit repeatedly and plot all paths; choose transformations that make paths simple, often linear on a log scale.
  3. Fit a random-coefficient (mixed-effects) path model by maximum likelihood, or for a quick look fit each unit separately.
  4. Compute the probability of crossing the critical level by each time by simulating paths from the fitted model; with the quick method, extrapolate each unit to the threshold and treat these pseudo failure times as data.
  5. Obtain bootstrap confidence intervals for the resulting distribution and its percentiles.
  6. Where some units have actually failed, compare with a censored time-to-event analysis of the same units.

Sources

  • Statistical Methods for Reliability Data (Meeker & Escobar, 1998)

Also known as

  • Degradation data analysis
  • Soft-failure analysis
  • Pseudo-failure-time method
  • General degradation path model
  • Degradation path analysis: estimating time-to-threshold distributions from repeated measurements of decline

Categories

  • Analytics & Behaviour
  • Statistics & Modelling