
Data collected in many fields often take the form of intervals due to measurement uncertainty, aggregation, or data protection needs. Conventional methods for analyzing data from designed experiments focus on single-valued observations and cannot be directly applicable to interval-valued data. Recent studies and critiques have increasingly focused on neutrosophic and simulation-based approaches. The former extend classical point-based statistical analysis through indeterminacy representations, while the latter preserve conventional inferential frameworks by sampling within observed intervals according to distributional assumptions. However, neutrosophic methods may neglect dependence structures among statistical quantities, whereas simulation-based methods depend on within-interval distributional assumptions that are often unknown or debated. This study addresses these challenges by proposing an optimization-based framework that derives feasible bounds for classical test statistics without imposing within-interval probability models. For completely randomized designs, the F-bounds are obtained by reformulating interval ANOVA as a box-constrained quadratic fractional programming problem, thereby extending classical ANOVA to interval-valued significance testing while preserving its inferential structure. The effectiveness of the proposed method is validated through both simulation studies and a real-life case study. This research is among the very first to provide a rigorous extension of classical ANOVA for interval settings. It bridges mathematical programming techniques with statistical hypothesis testing while ensuring globally optimal solutions. Notably, the method naturally reduces to classical ANOVA when interval data collapse to point values, providing a foundation for further methodological advancements in this field.
This paper introduces a class of Max-type nonparametric exponentially weighted moving average (Max-EWMA) schemes for high-dimensional processes based on interpoint and various Minkowski distances. The Max-EWMA monitoring scheme uses a plotting statistic based on the maximum of the absolute values of two EWMA statistics based on rank scores, one for the location and the other for the scale aspects of the distance distribution, equivalent to operating two EWMA schemes with standard control limits. The advantages of the Max-EWMA-type schemes over some existing high-dimensional Phase-II monitoring schemes are investigated. More specifically, several Monte Carlo simulations are conducted. The main conclusions are that the proposed Max-EWMA schemes exhibit robust in-control (IC) performance across different continuous multivariate distributions and, in the out-of-control (OOC) case, no distance metric completely outperforms the others. The Manhattan distance seems preferable when there are linear location shifts across all variables. In some sparse location-shift settings, Euclidean distance performs better for linear location shifts in an asymmetric multivariate population. In contrast, Chebyshev's distance performs better for a couple of symmetric multivariate densities. Finally, the implementation of the schemes is shown by monitoring production process data from a semiconductor manufacturing process.
Modern quality engineering regularly relies on high-dimensional data to evaluate complex product characteristics, process capability, and customer preferences. In these applications, data variation is typically governed by a lower-dimensional set of underlying latent features, a subset of which may be known a priori. Conditional multidimensional scaling performs dimensionality reduction while conditioning on these known features. The state-of-the-art conditional multidimensional scaling algorithm can accommodate missing values within the conditioning data, but with a computational complexity of up to O(N3). To address this bottleneck, we present a novel conditional multidimensional scaling framework for incomplete conditioning data that operates with O(N2) time complexity. We establish key theoretical properties of the proposed method and evaluate its practical utility. Empirical results, including a quality function deployment study of car perception and a part-to-part variation analysis for automotive head gasket manufacturing, demonstrate that our approach substantially reduces computational time while maintaining comparable performance in both dimensionality reduction and missing data imputation.
Traditional control charts such as Shewhart, EWMA, and CUSUM have limitations in simultaneously detecting both small and large process shifts efficiently. This article introduces an ensemble EWMA (E-EWMA) chart that addresses these issues by combining several EWMA components into a unified statistic with a single decision threshold, simplifying practical implementation and interpretation. The proposed chart is developed for Phase II monitoring under standard in-control parameter assumptions, and it applies to either individual observations or subgroup summaries (such as X & strns; ) when observations are independent over time. The numerical comparisons based on average run length, conditional expected delay, immediate signal proportion and weighted shift regime objectives show that the E-EWMA chart can improve the detection tradeoff when several separated shift regimes are important. The results also show that a single optimized EWMA remains preferable in some small shift dominated settings, which clarifies when the additional ensemble structure is warranted. The article also details practical design considerations, including robust optimization of smoothing parameters and component weights via a gradient-boosted surrogate model and differential evolution algorithm. Further, the impact of Phase 1 parameter estimation uncertainty on chart reliability is evaluated, emphasizing recalibration strategies to maintain performance. A custom-built interactive web application is provided that guides practitioners step-by-step through the design and optimization of E-EWMA chart. Application to real-world data from a screw-driving process illustrates the practical value of the E-EWMA approach, confirming its effectiveness in rapidly identifying shifts of varying magnitudes.
Weibull models are a central tool for reliability assessment, where engineers often need accurate confidence intervals for shape, scale and low lifetime quantiles under small samples, censoring and accelerated life testing. Standard likelihood-based intervals rely only on first-order asymptotics and can show substantial undercoverage, while existing refined methods are either tabulated for very specific settings or difficult to extend to more complex designs. We develop a unified higher-order likelihood framework for interval estimation in Weibull and log-Weibull models that treats any scalar parameter or reliability characteristic, including use-stress quantiles in accelerated tests, as the parameter of interest. Using the modified signed likelihood root together with an analytically constructed canonical parameter, we obtain simple closed-form adjustment factors that require no external tables and are valid for complete data, Type II censoring and accelerated life testing with log-linear scale models. Simulation studies across a wide range of shapes, sample sizes and designs show that the proposed intervals achieve coverage very close to the nominal level. A case study on breakdown times of an insulating fluid under accelerated testing illustrates the practical gains in estimating low use-stress quantiles.