A new calibration metric that considers statistical correlation: Marginal Probability and Correlation Residuals. (March 2020)

• Record Type: Journal Article
• Authors: Kim, Wongon; Yoon, Heonjun; Lee, Guesuk; Kim, Taejin; Youn, Byeng D.
• Is Part Of: Reliability engineering & system safety. Volume 195(2020)
• Publication Date: 2020-03
• Subjects: Optimization-based statistical model calibration -- Calibration metric -- Statistical correlation -- Marginal probability and correlation residual (MPCR); Reliability (Engineering) -- Periodicals; System safety -- Periodicals; Industrial safety -- Periodicals; Fiabilité -- Périodiques; Sécurité des systèmes -- Périodiques; Sécurité du travail -- Périodiques; 620.00452
• DOI: 10.1016/j.ress.2019.106677
• Languages: English
• ISSNs: 0951-8320

Abstract

Highlights: Neglecting statistical correlation leads to inaccurate statistical model calibration. The Marginal Probability and Correlation Residuals (MPCR) was newly proposed. The MPCR allows consideration of the statistical correlation effectively. The MPCR has normalization, boundedness, and marginalization. Abstract: Computer-aided engineering (CAE) models have been indispensable to virtual testing for designing and evaluating engineered systems to satisfy reliability requirements. However, it is not easy to fully characterize the variability in the model input variables due to limited resources. Statistical model calibration is thus of great importance as a strategy to improve the predictive capability of a CAE model. Optimization-based statistical model calibration is formulated as an unconstrained optimization problem that infers the unknown statistical parameters of input variables associated with a CAE model by maximizing statistical similarity between predicted and observed output responses . A calibration metric is defined as the objective function to be maximized that quantifies statistical similarity. One important challenge in formulating a calibration metric is how to properly consider the statistical correlation in output responses. Thus, this study proposes a new calibration metric: The Marginal Probability and Correlation Residual (MPCR) . The foundational idea of the MPCR is to decompose a multivariate joint probability distribution into multiple marginal probability distributions, while considering the statistical correlation between output responses. The MPCR has favorable properties, such as normalization, boundedness, and marginalization. Two mathematical and two engineering examples are presented to demonstrate the effectiveness and potential benefits of the MPCR.