Study On Loess Collapsibility Based On Data Mining

Collapsibility is one of the most important loessal behaviors as well as a baffling problem in loessal engineering research. Through loessal collapsibility engineering tests and the collection of a large bulk of data about loessal engineering, this thesis deals in depth with the factors affecting loessal collapsibility and the problems of collapsibility prediction based on data mining technology. The following research results were obtained:(1) The thesis suggests a new concept - using data mining technology to study loessal collapsibility. A methodological system is initially established with data mining technology to study loessal collapsibility so as to enrich the theory of loess collapsibility research, and provide a new kind of technology for loessal collapsibility research. The thesis also studies the analysis method of principal component analysis, CART decision trees, BP neural network, least-square supporting vector machine and particle swarm optimization in data mining technology. The BP artificial neural network is improved. At the same time, the thesis advances the feature selection algorithm and the correlative mining algorithm based on Gini coefficients (in the CART algorithm) and the half algorithm for reducing accuracy frequencies in the feature selection algorithm, as well as the method which uses the particle swarm optimization algorithm to optimize the parameters in the least square supporting vector machine. On the basis of this, the thesis studies and develops a data mining system of loessal collapsibility, and this system can be integrated with applied geo-tech software to carry out attraction and selection of features, correlation mining, classification and prediction mining. This system is important in maintaining and managing the database, thereby realizing the management of mining results, i.e. knowledge management and prediction factor management.(2) The feature selection technology is used to carry out quantity analysis of the effects of indices such as apparent resistivity, wave velocity, thermal conductivity, and heat capacity upon collapsibility. The results indicate that the apparent resistivity combined with wave velocity is the main factor affecting the collapsibility, and that the buried depth of the collapsible soil layer reflects soil stress and can be used as a prediction variable. The prediction model for loessal collapsibility is established using the improved particleswarm least square supporting vector machine algorithm. The experimental results from the model indicate that selection of prediction variables is reasonable, and the method of predicting collapsibility with apparent resistivity, wave velocity, and depth of collapsible soil layers is feasible.(3) The thesis suggests the concept of compaction rate via special tests and analysis of data for loessal collapsibility and its compaction (pressure) features. A new viewpoint is suggested that collapsibility can be evaluated with deformation property of disturbed loess samples. The correlated mining results indicate that the compaction rate is closely correlated with collapsibility, but remarkably in negative correlation with the collapsibility coefficients. In the compaction process, the closer the soil water content is to the optimal water content, the stronger this correlation is. The correlated mining results are analysed with soil structure theory.(4) Using conventional physical indices such as water content, dry density,liquid limitation,plastic limitation, and compressed coefficients, the principal component analysis method is used to carry out collapsibility classification and prediction mining of data compression, elimination of redundant information, and selection of prediction variables. The classification rules for the loessal collapsibility and non-collapsibility, the self-weight collapsibility and self-weight non-collapsibility are mined. As a result, the corresponding decision tree classification model is established. The model is tested using data from 2766 groups in 59 projects, and the testing results indicate that the classification model is of high accuracy. Also, the neural network model for predicting collapsibility coefficients, and at the same time, the particle swarm optimization least square supporting vector machine model for predicting the collapsibility coefficients of the variable prediction, are established on the basis of data for apparent resistivity, shear wave velocity and buried depth of soil layer. The practical engineering tests indicate that all the models are both accurate and practical.This paper was supported by the National Natural Science Foundation of China(No. 10572090) and Natural Science Foundation of Xi'an Design & Research Institute of Coal Industry,...