Robust Estimation in Step-Stress Experiments under Weibull Lifetime Distributions

1. Problem Statement

Modern industrial products (e.g., electronic components, batteries, aerospace parts) exhibit high reliability, leading to long lifetimes. Conducting standard life tests under normal operating conditions is often prohibitively time-consuming and costly. Accelerated Life Tests (ALTs), specifically Step-Stress Accelerated Life Tests (SSALTs), are used to induce failures earlier by increasing stress levels (e.g., temperature, voltage) at predetermined intervals.

However, traditional estimation methods, primarily the Maximum Likelihood Estimator (MLE), suffer from a critical weakness: they are highly sensitive to data contamination and outliers. In real-world scenarios, data may contain measurement errors or anomalous failures that can severely bias MLE results, rendering them unreliable. While robust estimation methods exist, there is a gap in applying them to SSALT models with continuous monitoring under Weibull lifetime distributions, particularly when dealing with mixed data types (exact failure times and right-censored observations).

2. Methodology

The authors propose a robust estimation framework based on Minimum Density Power Divergence Estimators (MDPDEs).

• Model Setup:

  • Experiment: A simple SSALT with two stress levels ($x_1, x_2$) and Type-I censoring (experiment ends at a fixed time $\tau_2$).
  • Distribution: Weibull distribution $W(\lambda, \eta)$ with scale parameter $\lambda$ and shape parameter $\eta$.
  • Stress Relationship: The scale parameter follows a log-linear relationship with stress: $\lambda_i = \exp(a_0 + a_1 x_i)$. The shape parameter $\eta$ is assumed constant across stress levels.
  • Cumulative Exposure Model (CEM): Used to model the transition between stress levels, ensuring the cumulative distribution function (CDF) remains continuous.
  • Data Structure: The observed data follows a mixed distribution: a continuous component for exact failure times occurring before $\tau_2$, and a discrete point mass at $\tau_2$ for right-censored units (survivors).

• Estimation Approach:

  • MDPDE Framework: Instead of maximizing the likelihood, the method minimizes the Density Power Divergence (DPD) between the empirical distribution of the data and the parametric model.
  • Tuning Parameter ($\beta$): The DPD includes a tuning parameter $\beta \geq 0$.
    • When $\beta = 0$, the MDPDE reduces to the standard MLE.
    • As $\beta$ increases, the estimator becomes more robust to outliers, trading off some efficiency for resistance to contamination.
  • Derivation: The authors derive explicit closed-form expressions for the objective function $H_n^\beta$ by integrating the DPD over the mixed distribution (continuous intervals and the censoring point). This involves using the Incomplete Gamma function to handle the integrals of the Weibull density raised to the power of $\beta+1$.

• Asymptotic Theory:

  • The paper establishes the asymptotic normality of the MDPDE vector $\hat{\theta}_\beta = (\hat{a}_0, \hat{a}_1, \hat{\eta})$.
  • Explicit formulas are derived for the asymptotic variance-covariance matrix $\Sigma = J_\beta^{-1} K_\beta J_\beta^{-1}$, where $J_\beta$ is the Jacobian of the DPD loss and $K_\beta$ relates to the second moment of the score function.
  • The Delta Method is applied to derive asymptotic distributions for derived quantities of interest, such as Mean Time to Failure (MTTF), Reliability functions, and Quantiles.

3. Key Contributions

1. Extension to Weibull SSALT: The paper extends the MDPDE methodology, previously applied to exponential lifetimes or interval-censored data, to the more flexible Weibull distribution under continuous monitoring (exact failure times).
2. Handling Mixed Distributions: It provides a rigorous theoretical derivation for MDPDEs in the context of mixed continuous-discrete data arising from Type-I censored step-stress experiments.
3. Explicit Asymptotic Formulas: The authors provide detailed, explicit mathematical expressions for the elements of the Jacobian ($J_\beta$) and the vector $\xi_\beta$, which are necessary for constructing confidence intervals and hypothesis tests.
4. Robust Inference for Derived Metrics: The framework allows for the robust estimation of practical reliability metrics (MTTF, reliability at time $t$, quantiles) and the construction of robust confidence intervals using logit or log transformations to handle boundary constraints.

4. Results

The performance of the proposed MDPDEs was evaluated through extensive simulation studies and a real-data example.

• Simulation Study:

  • Setup: Data was generated from a Weibull SSALT model and contaminated with outliers (up to 10%) by altering specific model parameters ($a_0, a_1, \eta$).
  • Findings:
    • Efficiency vs. Robustness: In the absence of outliers, MLE ($\beta=0$) performed best. However, even small contamination levels (3%) caused MLE estimates to deviate significantly.
    • Tuning Parameter Effect: Larger values of $\beta$ (e.g., 0.6–1.0) maintained high accuracy and low bias in the presence of outliers, effectively mitigating the influence of contaminated data.
    • Confidence Intervals: MLE-based confidence intervals suffered from severe under-coverage (dropping to <20% in some cases) under contamination. MDPDE-based intervals with higher $\beta$ maintained coverage probabilities close to the nominal 95% level.
    • Bias Reduction: The MDPDE effectively reduced the bias in parameter estimates and derived metrics (like MTTF) caused by outliers.

• Real-Data Example:

  • Data: A dataset involving solar lightning devices tested under two temperature stress levels.
  • Analysis: The MDPDEs were applied with varying $\beta$. The results showed that while point estimates were relatively stable (suggesting the real data may not have extreme outliers), the confidence intervals widened as $\beta$ increased. This demonstrates the method's ability to provide a "safety margin" in uncertainty quantification when the data quality is uncertain.
  • Insights: The analysis confirmed that increased temperature significantly reduced the Mean Time to Failure (MTTF) and reliability of the devices.

5. Significance

This paper makes a significant contribution to the field of reliability engineering and statistical inference:

• Practical Robustness: It offers a practical solution for industries where data contamination is a risk but cannot be easily identified or removed. The MDPDE provides a "safe" estimation alternative that does not collapse under outlier influence.
• Theoretical Rigor: By deriving the asymptotic properties for the Weibull SSALT model, the paper enables the construction of valid statistical tests (Wald, Rao) and confidence intervals, which were previously unavailable for robust estimators in this specific setting.
• Flexibility: The methodology is adaptable. The authors suggest future work could extend this to other distributions (Log-normal, Gamma) and censoring schemes (Type-II, interval-censored), broadening the applicability of robust life-testing methodologies.
• Decision Support: By providing reliable estimates of MTTF and reliability even with noisy data, the method supports better decision-making in product design, warranty planning, and maintenance scheduling.

In summary, the paper successfully bridges the gap between robust statistical theory and the practical complexities of step-stress accelerated life testing, offering a superior alternative to classical MLE when data integrity is compromised.