Robust Optimum Life-Testing Plans under Progressive Type-I Interval Censoring Schemes with Cost Constraint

Xiao-Dong Zhou, Yunjuan Wang, Rong-Xian YueSymmetry 2022 · DOI

TL;DR

This paper introduces robust methods for designing optimal life-testing plans when data is progressively interval-censored and model parameters are uncertain. By developing Bayesian and minimax compound design criteria that account for both parameter estimation and quantile estimation, along with cost constraints, the authors provide practical tools for engineers to create more reliable and efficient life-testing experiments.

What Problem Does This Paper Address?

Designing effective life-testing experiments is crucial for understanding product reliability. However, real-world life tests often involve "progressive censoring," where units are removed from the test before failure, and "interval censoring," where failures are only observed within specific time intervals. These schemes save time and cost but complicate data analysis. A key challenge is that optimal experimental designs typically depend on assumed values for the underlying statistical model parameters (e.g., for the Weibull distribution). If these assumed values are incorrect, the resulting "locally optimal" designs may perform poorly. Furthermore, experimenters often have multiple objectives, such as accurately estimating model parameters and estimating specific quantiles of the lifetime distribution. This paper addresses these issues by proposing robust design strategies that are less sensitive to parameter uncertainty and can balance multiple design objectives, all while considering practical cost constraints.

Methods

The authors propose two main robust design strategies:

  • Bayesian Compound Design: This approach incorporates prior knowledge about the model parameters (Weibull distribution parameters $\mu$ and $\sigma$) by using probability distributions (censored normal and inverse Gamma distributions) to represent their uncertainty. It then optimizes a "compound" criterion that balances two objectives: D-optimality (for estimating model parameters) and c-optimality (for estimating quantiles). The Bayesian compound criterion is denoted as $\Psi_{BCD}(\xi; \kappa)$, where $\kappa$ is a weight reflecting the relative importance of the two objectives.
  • Minimax Compound Design: This strategy aims to minimize the maximum possible loss (or maximize the minimum performance) over the range of uncertain parameter values. It also uses a compound criterion, $\Psi_{McD}(\xi; \omega)$, balancing D- and c-optimality, where $\omega$ is a weight parameter.

To handle the practical aspect of limited budgets, the paper incorporates cost constraints into the design optimization. Two algorithms are presented to find these robust optimal designs:

  • Mixed-Integer Nonlinear Optimization (MNO) Algorithm: This algorithm is used for a simplified case where inspection intervals are of equal length and removal proportions are constant.
  • Particle Swarm Optimization (PSO) Algorithm: This more general algorithm is employed to find optimal designs for the full progressive type-I interval censoring scheme, allowing for variable inspection times and removal proportions.

Datasets

The paper uses two types of datasets for evaluation:

  • Simulated Data: Data is generated based on Aggarwala's [4] method for a Weibull distribution with specified parameters and removal proportions. This allows for controlled testing of the proposed methods.
  • Real-Life Data: The authors analyze a dataset of 112 patients with plasma cell myeloma treated at the National Cancer Institute, previously studied in other reliability analyses. This demonstrates the practical applicability of their methods to real-world scenarios.

Key Insights and Ideas

  • Robustness is Crucial: Designs that are optimal only for specific "planning values" of parameters can be highly suboptimal if those values are slightly off. Robust designs (Bayesian and minimax) offer a significant improvement in performance across a range of parameter values.
  • Balancing Objectives: The compound design criteria (Bayesian and minimax) are essential for experimenters who have multiple goals, such as estimating both model parameters and specific quantiles (e.g., median lifetime). The weight parameters ($\kappa$ and $\omega$) allow for flexible trade-offs between these objectives.
  • Cost Matters: Incorporating cost constraints is vital for practical experimental design. The proposed algorithms effectively find optimal plans within a given budget, demonstrating that efficiency can be achieved even with resource limitations.
  • Removal Proportions Impact Design: The rate at which units are removed from the test significantly influences the optimal number of inspections and their timing. Higher removal rates generally lead to fewer inspections but longer intervals between them.
  • Uncertainty Increases Complexity: As the uncertainty in model parameters increases, the optimal designs tend to require more inspection times and longer experimental durations to maintain robustness.

Key Results

  • The paper demonstrates that the proposed Bayesian and minimax compound designs consistently outperform locally optimal designs, especially when parameter uncertainty is considered.
  • The efficiencies of the robust designs (BcD- and McD-optimal) relative to locally optimal designs are shown to be high across various scenarios.
  • Numerical examples show that the number of inspections and the duration of the experiment are influenced by cost parameters, removal proportions, and the degree of parameter uncertainty. For instance, increasing the total budget ($C$) generally leads to more test units ($N$), while increasing the cost per unit ($C_s$) reduces $N$.
  • The study provides tables of optimal test plans (number of units $N$, number of inspections $k$, inspection times $t_j$, and interval lengths $\tau$) for various scenarios, allowing practitioners to select appropriate plans based on their specific needs and constraints.

Limitations and Future Work

  • Approximation Accuracy: The Bayesian and minimax criteria are approximated using a finite set of percentiles from prior distributions. The authors suggest that using more advanced Monte Carlo sampling methods or advanced optimization techniques could further improve the accuracy of these approximations.
  • Distributional Uncertainty: The current work assumes the Weibull distribution. Future research could explore robust designs for situations where the underlying lifetime distribution itself is uncertain or when dealing with other distributions (e.g., lognormal, generalized exponential).
  • Other Uncertainties: The paper focuses on parameter uncertainty. Future work could investigate robust designs that account for other sources of uncertainty, such as measurement errors or variability in censoring times.
  • Extension to Other Schemes: The proposed methods are designed for Progressive Type-I Interval Censoring (PIC-I). The authors suggest that these approaches can be extended to other censoring schemes, such as accelerated life tests or degradation tests.

Summary updated

BibTeX
@article{zhou2022robust,
  author = {Xiao-Dong Zhou and Yunjuan Wang and Rong-Xian Yue},
  title = {{Robust Optimum Life-Testing Plans under Progressive Type-I Interval Censoring Schemes with Cost Constraint}},
  journal = {Symmetry},
  year = {2022},
  doi = {10.3390/sym14051047},
}
Figure