The Research Of Slope Stability Analysis Based On Machine Learning

Slope stability analysis is one of the traditional subjects in geotechnical engineering.In the 1980 s,machine learning theory was introduced into this field,which simplified the analysis process by constructing the mapping relationship between slope stability and influence factors.But for a long time,the error of correlation method is large.In recent years,machine learning theory has made breakthroughs,and its accuracy in biometrics,text processing and other applications has been greatly improved.However,the machine learning method of slope stability analysis has not been followed up,and the research is relatively lagging behind,especially the rough design of learning algorithm,which needs to be improved.This paper traces the development of machine learning applied to slope stability research,and discusses the application process,important concepts and methods of machine learning.Through the collection of slope sample data in the literature,the data sets with 72 slopes labeled with safety factor and 148 slopes labeled with stable state are established.According to the relevant theory,the learning algorithm is designed,especially the elements of the learning algorithm are selected and analyzed by using grid search and random search methods.The feedforward neural network for estimating the safety factor of slope and judging the stability state of slope is established,and different data are used for testing and comparison.The program is written based on Python programming language,Keras deep learning library and scikit-learn machine learning library.There are the following conclusions:(1)After using grid search method to select the elements of the learning algorithm,the accuracy and stability of the feedforward neural network model for safety factor estimation will be greatly improved,which has practical application value.(2)The feedforward neural network model for judging stable state uses the original data to train and has high accuracy.The model can reduce the workload of slope stability state investigation and ensure the correctness of the result of instability.(3)For the feedforward neural network to estimate the safety factor,the learning algorithm generally uses the optimization algorithm adadelta to perform the best,Adam second,and adagrad worst;setting the bias to 0.1 has some advantages over 0;Lecun's initialization strategy is better,and whether the initial parameters are uniform or normal distribution has no effect.For the feedforward neural network to judge the stable state,tanh has a significant advantage,relu is the second,and logistic is the worst.(4)Feed-forward neural network has advantages over support vector machine.Expanding the scale of data set and the selection range of learning algorithm elements,collecting data set scientifically and reasonably can improve the performance of neural network.Compared with grid search,random search has less manual participation,wider range of super parameter selection and more efficient.