Slope reliability analysis using Bayesian optimized convolutional neural networks. Issue 8 (16th August 2022)

• Record Type: Journal Article
• Authors: Lin, Houlai; Li, Liang; Meng, Kaiqi; Li, Chunli; Xu, Liang; Liu, Zhiliang; Lu, Shibao
• Is Part Of: Engineering computations. Volume 39:Issue 8(2022)
• Page Start: 3012
• Page End: 3037
• Publication Date: 2022-08-16
• Subjects: Slope stability -- Spatial variability -- Bayesian optimization -- Convolutional neural networks -- Artificial neural networks -- Response surface method; Computer-aided engineering -- Periodicals; Computer graphics -- Periodicals; 620.00285
• DOI: 10.1108/EC-01-2022-0026
• Languages: English
• ISSNs: 0264-4401

Abstract

Abstract : Purpose: This paper aims to develop an effective framework which combines Bayesian optimized convolutional neural networks (BOCNN) with Monte Carlo simulation for slope reliability analysis. Design/methodology/approach: The Bayesian optimization technique is firstly used to find the optimal structure of CNN based on the empirical CNN model established in a trial and error manner. The proposed methodology is illustrated through a two-layered soil slope and a cohesive slope with spatially variable soils at different scales of fluctuation. Findings: The size of training data suite, T, has a significant influence on the performance of trained CNN. In general, a trained CNN with larger T tends to have higher coefficient of determination ( R 2 ) and smaller root mean square error (RMSE). The artificial neural networks (ANN) and response surface method (RSM) can provide comparable results to CNN models for the slope reliability where only two random variables are involved whereas a significant discrepancy between the slope failure probability ( P f ) by RSM and that predicted by CNN has been observed for slope with spatially variable soils. The RSM cannot fully capture the complicated relationship between the factor of safety (FS) and spatially variable soils in an effective and efficient manner. The trained CNN at a smaller the scale of fluctuation ( λ ) exhibits a fairly good performance in predicting the P f for spatially variable soils at higher λ with a maximum percentage error not more than 10%. The BOCNN has a larger R 2 and a smaller RMSE than empirical CNN and it can provide results fairly equivalent to a direct Monte Carlo Simulation and therefore serves a promising tool for slope reliability analysis within spatially variable soils. Practical implications: A geotechnical engineer could use the proposed method to perform slope reliability analysis. Originality/value: Slope reliability can be efficiently and accurately analyzed by the proposed framework.