Neural network ensemble-based sensitivity analysis in structural engineering: Comparison of selected methods and the influence of statistical correlation. (1st January 2021)

• Record Type: Journal Article
• Authors: Pan, Lixia; Novák, Lukáš; Lehký, David; Novák, Drahomír; Cao, Maosen
• Is Part Of: Computers & structures. Volume 242(2020)
• Publication Date: 2021-01-01
• Subjects: Global sensitivity analysis -- Local sensitivity -- Statistical correlation -- Neural network ensemble; Structural engineering -- Data processing -- Periodicals; Electronic data processing -- Structures, Theory of -- Periodicals; 624.171
• DOI: 10.1016/j.compstruc.2020.106376
• Languages: English
• ISSNs: 0045-7949

Abstract

Highlights: The NNE is testified a superior surrogate model for parameter sensitivity analysis. Properties of typical methods of parameter sensitivity analysis are fully clarified. Strategy for sensitivity analysis of correlated variables are elaborated. Mechanism of local and global methods for sensitivity analysis is revealed. Robustness & accuracy of NNE in sensitivity analysis are demonstrated. Abstract: Surrogate model-based sensitivity analysis, especially framed by neural network ensemble (NNE), is an attractive but unresolved issue in structural reliability assessment. In this paper, differing from existing studies, an overview and assessment of typical methods for surrogate model-based parameter sensitivity analysis, namely the input perturbation method, the local analysis of variance, the connection weight method, the non-parametric Spearman rank-order correlation method, and the Sobol indices method, are performed and demonstrated on three illustrative cases of increasing complexity: a simple theoretical instance, an engineering case of midspan deflection of a simply-supported beam, and a real-world practical application of shear failing in a precast concrete girder. Through comprehensive comparisons, several findings are obtained as follows: (i) the NNE is testified a superior surrogate model for sensitivity analysis to a single artificial neural network; (ii) robustness and accuracy of an NNE in sensitivity analysis are demonstrated; (iii) the properties of these parameter sensitivity analysis methods are fully clarified with distinguished merits and limitations; (iv) mechanism of local- and global- sensitivity analysis methods is revealed; and (v) the strategy for sensitivity analysis of correlated descriptive variables are elaborated to address the impact of correlation among random variables in engineering systems.