Slope failure is a significant concern in geotechnical engineering and can compromise soil stability. It can occur for various reasons, including weak soil conditions, heavy rainfall that washes away surface materials, seismic activity, overloading, and deforestation. Analyzing the failure mechanisms is crucial, and it predominantly relies on the limit equilibrium method.
However, traditional methods often do not account for various underlying factors, such as geological materials. These uncertainties pose challenges that classic theories struggle to address effectively. This
situation makes the application of conventional estimation procedures quite difficult. Consequently, risk analysis and assessment have emerged as essential tools for managing the uncertainties associated with slope failures. It is important to note that not all uncertainties are random; some can be objectively
quantified, particularly those arising from incomplete information.
To address the challenges associated with non-linear relationships, artificial neural networks (ANN) have been extensively utilized by researchers across various fields. Their application is particularly notable
in geotechnical issues, including rock engineering, pavement analysis, landslide susceptibility, and soil liquefaction. Noteworthy in this context is the work conducted by Tsung-lin Lee et al., from the Department of Construction and Facility Management at Leader University in Tainan, China. They employed ANN in their research focused on assessing slope failure along the Cross-Island Highway in Taiwan. This highway, constructed in 1973, was not opened to the public until 1993 due to recurrent landslide incidents.
This research study utilized a dataset of 340 samples collected from in-situ surveys conducted along the South Cross-Island Highway. Eight distinct training and validation time series, identified as Cases A1 through A8, were analyzed throughout the study. Each case incorporated varying numbers of failure and non-failure data sets. For instance, the Artificial Neural Network (ANN) training for Case A1 consisted of 15 “failure” data points and 85 “non-failure” data points, while validation comprised 11 “failure” and 29 “non-failure” data series. The analysis employed eight input parameters, which included slope gradient angle, slope height, daily rainfall, cumulative precipitation, surface acceleration, material strength, slope direction, and earthquake magnitudes.
The hyperparameters used by the artificial neural network (ANN) were determined through trial and error. These hyperparameters include the number of hidden units set to 7, a learning rate (η) of 0.01, a momentum factor (α) of 0.9, and a specific number of training iterations (epochs=17000). The authors developed eight models (B1-B8) by considering seven input parameters, excluding one parameter in each model.
The results indicate that omitting the vegetation condition slightly reduces the success ratio to 87.5%. Other parameters, such as the number of joints and slope direction, show low contributions at 2.5% and 5.0% respectively, which leads to minor decreases in the success ratio. Despite these reductions in input data, the ANN maintains a satisfactory success ratio of 80.0%. This suggests that while using all parameters could enhance predictions of slope stability, the availability of eight observed parameters—including slope gradient angle, slope height, cumulative precipitation, daily rainfall, material strength, joint number, vegetation condition, and slope direction—produces a success ratio of 95.0% in neural simulations of slope failure.
The case study discussed compellingly demonstrated that the neural methodology employed can be effectively utilized for estimating slope failures. This finding underscores the importance of thorough training and validation of the methodology using data sets specific to the site of interest before applying it to different geographical areas or under varying environmental conditions. Such an approach ensures the accuracy and reliability of the predictions made by the neural network, allowing for better-informed decision-making in slope management and risk assessment.
Reference
Lee, Tl., Lin, Hm. & Lu, Yp. Assessment of highway slope failure using neural networks. J. Zhejiang Univ. Sci. A 10, 101–108 (2009). https://doi.org/10.1631/jzus.A0820265