Global Sensitivity Analysis Using Space Partition Technique

Recently,sensitivity analysis has been becoming one of the most attracting research branch in the field of random uncertainty analysis for structures and systems.More and more researchers are paying much attention to this direction,especially to the global sensitivity analysis.The importance rank of input variables,which affect the uncertainty of outputs and risk level,can be obtained through the global sensitivity analysis methods,and this rank will also provide guidance on optimization design,performance improvement,and etc.Many methods have been developed for the calculation of various global sensitivity indices,however,how to deal with the high dimensional issue and apply them on the complex structures and systems are still very challenging.The sampling-based methods are widely used in the global sensitivity analysis,but the high computation cost and low sample-usage-rate of the existing sampling-based methods have limited their application for the high dimensional and complicated engineering structures.In order to overcome the shortcomings of the existing methods,space partition technique is presented to estimate the various sensitivity indices in the global sensitivity analysis.It can greatly increase the utilization of the available samples and improve the computational efficiency and practicability.The first mainly work of this study is combing space partition technique and Monte Carlo simulation to calculate the variance-based sensitivity index.Three practical formulations are derived in this study,and their accuracy and robustness are analyzed from both the theoretical and technical view.The second work of this study is introducing space partition technique into the cross entropy method and subset simulation method.The proposed methods can calculate the failure-probability-based sensitivity indices using the same samples which have been generated for estimating failure probability.Therefore,there is no additional function call required by the global sensitivity analysis,and the proposed methods take fully advantage of these efficient sampling-base methods for reliability analysis.