Multi-fidelity deep neural network with Monte Carlo dropout technique for uncertainty-aware risk recognition of backward erosion piping in dikes

Liu, Hongchen, Su, Huaizhi, Yang, Jiaquan and Li, Haijiang ORCID: https://orcid.org/0000-0001-6326-8133 2024. Multi-fidelity deep neural network with Monte Carlo dropout technique for uncertainty-aware risk recognition of backward erosion piping in dikes. Applied Soft Computing 166 , 112165. 10.1016/j.asoc.2024.112165 Item availability restricted.

PDF - Accepted Post-Print Version Restricted to Repository staff only until 28 August 2025 due to copyright restrictions. Download (3MB)

Abstract

Backward erosion piping (BEP) is an increasingly critical failure mechanism in dike systems, often triggered by floods resulting from extreme rainfall events, which are exacerbated by the ongoing shifts in climate patterns. This study embarks on a nuanced examination of BEP, a phenomenon that substantially undermines flood control and emergency management endeavors. Addressing the complexities of BEP incidents and the consequential variable impact on dike infrastructure, we have incorporated Monte Carlo dropout techniques within a multi-fidelity deep learning framework to enhance predictive accuracy and provide an uncertainty-aware assessment of BEP risk. Drawing on 16 sets of physical model experiments emulating dike structures from the lower Yangtze River basin in China, we conducted a detailed study of the BEP evolution mechanism under various hydrological scenarios intensified by climate change. These experiments allowed us to observe the initiation and progression of BEP in different conditions. The development of a structural equation model quantifies the effects of several critical factors—including those exacerbated by climatic variability—on BEP dynamics. Seven key factors were identified as influential to BEP risk levels, integrating distinct BEP traits, hydrological attributes, and dike engineering conditions. A Monte Carlo dropout-enhanced multi-fidelity deep neural network (MFDNN) was crafted, synthesizing low-fidelity experimental data with high-fidelity field case studies, to construct an advanced model for BEP risk level identification. Compared against four sophisticated machine learning models, our MFDNN demonstrated superior performance, effectively blending experimental insights with real-world occurrences. The proposed model emerges as a scientifically robust and pragmatic tool for delineating BEP risk levels in dike systems, providing vital guidance for categorizing BEP incidents and forging targeted, climate-informed emergency response strategies.

Item Type:	Article
Date Type:	Publication
Status:	Published
Schools:	Engineering
Publisher:	Elsevier
ISSN:	1568-4946
Date of First Compliant Deposit:	14 September 2024
Date of Acceptance:	22 August 2024
Last Modified:	07 Nov 2024 23:00
URI:	https://orca.cardiff.ac.uk/id/eprint/172106

Actions (repository staff only)

Edit Item