مقاله انگلیسی رایگان در مورد ارزیابی روانگرایی خاک با استفاده از ترکیب فناوری AI و مدل کوپل ENN / t-SNE – الزویر 2020

1- Introduction

2- Brief review of existing methods

3- Methodological approach

4- Case study: liquefaction potential assessment

5- Conclusions

References

Abstract

This paper presents a new evolutionary neural network (ENN) algorithm coupled with the dimensionality reduction technique ‘t-distributed stochastic neighbour embedding’ (t-SNE). The ENN model features the crossbreeding of a differential evolution method and a stochastic gradient optimisation algorithm. The t-SNE is used to visualise the training and testing datasets and the ENN model performance. The proposed ENN model is applied to a relatively large soil liquefaction database. The good convergence and generalisation ability of the proposed model and the negligible misclassification results demonstrate that the proposed ENN model can provide accurate, efficient, and flexible results. The prominent and practical abilities of t-SNE to recover the structure of the initial conditions and to demonstrate the ENN model performance are discussed. This coupled approach simplifies the analysis and/or prediction of hazards for which large quantities of data are required.

Introduction

One of the focuses of civil engineers in the past three decades has been to tackle the soil liquefaction phenomenon efficiently, given the random nature of its occurrence. Soil liquefaction occurs when a saturated or partially saturated soil substantially loses strength and stiffness in response to an applied stress (such as shaking during an earthquake or other sudden changes in its stress condition), in which a material that is ordinarily a solid behaves like a liquid [1,2]. Soil liquefaction occurs when the effective stress (shear strength) of the soil is reduced to essentially zero and imposes a real risk on the surrounding structures [3–8]. The dedication of engineers and researchers has thus been triggered and further reinforced by a plethora of casualties [9] and the severity of structural damages [10–13] caused by this phenomenon. However, the prediction of soil liquefaction has remained a formidable task owing to the nonlinearity of soil behaviour [14–22] and the characteristics of the seismic power dissipation. The traditional methods for predicting soil liquefaction [23–26] still have numerous limitations, which have initiated the proposals of more powerful (efficient, flexible, and accurate) strategies. Among them, data mining (DM) has attracted particular attention recently. Data mining (DM) is a multidisciplinary subfield of computer science that includes statistics, database technology, and machine learning for the analysis of previously unknown or unsuspected relationships buried in large datasets. The DM algorithms are increasingly accepted in geotechnical engineering and have generally shown an excellent ability to solve multi-dimensional and complex problems [27].