Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
IEC 61710: Power Law Model — Goodness-of-Fit Tests and Confidence Intervals
Tip: IEC 61710 provides the statistical foundation for modelling the reliability of repairable systems using the power law (Weibull process) model. It is an essential reference for reliability engineers working with failure time data from systems that undergo repairs or maintenance.
1. Scope and Mathematical Foundation
IEC 61710 specifies procedures for fitting a power law model to failure time data from repairable systems and for assessing the adequacy of the fit. The power law model, also known as the Weibull process or the Non-Homogeneous Poisson Process (NHPP) with power law intensity function, describes the instantaneous failure intensity as a function of time:
Power Law Intensity Function:
λ(t) = (β/θ) · (t/θ)β-1
Where:
λ(t) = instantaneous failure intensity at time t
β = shape parameter (dimensionless)
θ = scale parameter (time units)
The cumulative number of failures by time t is:
E[N(t)] = (t/θ)β
The shape parameter β governs the reliability trend of the system. When β = 1, the process reduces to a homogeneous Poisson process (HPP) with constant failure intensity, indicating no reliability growth or degradation. When β < 1, the failure intensity decreases over time, indicating reliability growth — a typical pattern observed during developmental testing programs. When β > 1, the failure intensity increases over time, indicating system degradation or wear-out. This simple yet powerful characterization makes the power law model one of the most widely used tools in reliability engineering for repairable systems.
Warning: A common mistake in reliability analysis is applying the power law model to non-repairable systems or to data where the “repair” does not restore the system to an operational state comparable to its condition before failure. The model assumes that repairs are “minimal” — restoring function without changing the overall failure intensity process.
The standard covers three primary applications: point estimation of model parameters, interval estimation (confidence intervals) for the shape parameter β and the scale parameter θ, and goodness-of-fit testing to determine whether the power law model adequately represents the observed failure data. Both maximum likelihood estimation (MLE) and graphical estimation methods are described, allowing engineers flexibility depending on data quality and available computational tools.
2. Statistical Procedures and Goodness-of-Fit Tests
IEC 61710 defines several goodness-of-fit tests to validate the power law model assumption. The choice of test depends on whether the failure data are time-truncated (test stopped at a predetermined time T) or failure-truncated (test stopped after a predetermined number of failures n).
| Test Type | Test Statistic | Truncation | Null Distribution |
|---|---|---|---|
| Cox–Lewis (modified) | Under HPP assumption, transformed interarrival times are exponential | Time-truncated | χ2 with 2n degrees of freedom |
| Military Handbook (MIL-HDBK-189) | CM = 2 ∑ ln(T/ti) | Time-truncated | χ2 with 2n degrees of freedom |
| Bartlett’s test (modified) | B = 2n [ln(∑ti/n) − ∑ln(ti)/n] / [1 + (n+1)/(6n)] | Failure-truncated | χ2 with n−1 degrees of freedom |
| Laplace trend test | U = (∑ti − nT/2) / (T √(n/12)) | Time-truncated | Standard normal N(0,1) |
The most widely used test in practice is the MIL-HDBK-189 test (often called the Crow test). The test statistic CM is compared against critical values from the chi-squared distribution. If the test statistic falls within the acceptance region, there is insufficient evidence to reject the power law model. A significant result indicates that the power law model may not adequately describe the failure process, and alternative models such as the log-linear process or the modulated power law should be considered.
The standard also provides detailed procedures for computing confidence intervals. For the shape parameter β, the confidence interval is based on the fact that 2nβ/βˆ (where βˆ is the MLE of β) follows a chi-squared distribution with 2(n−1) degrees of freedom for failure-truncated data. For time-truncated data, the degrees of freedom adjustment differs, reflecting the additional uncertainty from the stopping condition. These confidence intervals are essential for making engineering decisions about whether the observed reliability trend is statistically significant or merely due to random variation.
Good Practice: Always compute both point estimates and 90% or 95% confidence intervals for β. If the confidence interval includes 1.0, you cannot conclude that the system exhibits reliability growth or degradation — the data are consistent with a constant failure intensity (HPP). In such cases, consider extending the test duration or collecting additional failure data.
3. Engineering Design Insights and Practical Applications
The power law model under IEC 61710 finds broad application across multiple industries. In aerospace and defence, it is the standard model for tracking reliability growth during developmental testing, as codified in MIL-HDBK-189 and the Crow/AMSAA methodology. In the automotive industry, it is used to analyse warranty claims data where vehicles are repaired and returned to service. In industrial equipment maintenance, the model helps optimise preventive maintenance intervals by detecting the onset of wear-out (β > 1) early.
One of the most powerful engineering applications is the use of the power law model for reliability growth planning. By tracking the cumulative number of failures against cumulative test time on log-log axes, engineers can estimate the current reliability level and project future reliability under continued testing. The slope of the log-log plot is β, and the intercept provides information about the initial reliability level. A typical reliability growth program might set a target β value of 0.5 to 0.7, representing a significant reduction in failure intensity as testing progresses.
| β Range | Reliability Trend | Engineering Interpretation |
|---|---|---|
| β < 0.5 | Strong growth | Aggressive corrective actions; potential over-testing early |
| 0.5 ≤ β < 0.8 | Moderate growth | Effective FRACAS; healthy development program |
| 0.8 ≤ β < 1.0 | Weak growth | Mature system; diminishing returns from testing |
| β = 1.0 | No trend | Constant failure intensity; HPP applies |
| 1.0 < β < 1.5 | Mild degradation | Early wear-out; review preventive maintenance |
| β ≥ 1.5 | Strong degradation | Accelerating failure rate; redesign needed |
Critical: The power law model assumes that the failure process is a non-homogeneous Poisson process, which requires that the times between failures be conditionally independent given the intensity function. If failures cluster due to common-cause mechanisms (e.g., batch defects, environmental stresses), the model may produce misleading results. Always combine statistical tests with engineering judgement and failure mode analysis.
Practical implementation of IEC 61710 requires careful attention to data quality. Failure times must be recorded in continuous time (cumulative test time) rather than as discrete operating intervals. Right-censored data (systems that have not failed by the end of the observation period) can be accommodated, but the standard provides specific procedures for handling suspensions. For systems with multiple identical units under test, the standard recommends pooling the data assuming a common β but potentially different θ values, reflecting different starting reliabilities but common improvement rates.
Modern software tools have made the computational aspects of IEC 61710 straightforward, but the engineering interpretation remains challenging. A system that shows β = 0.6 with a narrow confidence interval of [0.55, 0.65] provides strong evidence of reliability growth and allows confident projection of future reliability. In contrast, β = 0.6 with a wide confidence interval of [0.35, 1.05] from limited data should be interpreted cautiously — the system may not be improving at all. The standard emphasises that confidence interval width is inversely proportional to the number of failures observed, providing a quantitative basis for planning test duration.
Frequently Asked Questions
Q1: What is the minimum number of failures required to apply IEC 61710?
The standard recommends at least 5 to 10 failures for meaningful parameter estimation. With fewer than 5 failures, confidence intervals become extremely wide, and goodness-of-fit tests lack statistical power to detect model inadequacy. For reliability growth tracking during developmental testing, a practical minimum is 20 failures to establish a reliable trend.
Q2: Can the power law model handle multiple systems with different test durations?
Yes. IEC 61710 provides procedures for combining data from multiple systems. The standard recommends fitting a common β across all systems while allowing individual θ parameters. This approach is appropriate when all systems undergo similar corrective action processes but start from different baseline reliability levels. The cumulative test time is the sum of individual system test times.
Q3: How does the power law model differ from the standard Weibull distribution?
The standard (2-parameter) Weibull distribution models the time-to-first-failure for non-repairable populations. The power law model (Weibull process) in IEC 61710 models the failure intensity over time for repairable systems. They share the same mathematical form but describe fundamentally different physical situations. Using a Weibull distribution on repairable system data is a common but serious statistical error.
Q4: What alternative models should be considered when the power law fit is inadequate?
When the goodness-of-fit test rejects the power law model, consider the log-linear model (exponential intensity function), the modulated power law model (also known as the Musa-Okumoto process), or the segmented power law model with different β values for different phases of testing. The choice depends on the nature of the departure from the power law assumption, as revealed by residual analysis and trend plots.
