Summary: | 碩士 === 國立雲林科技大學 === 營建工程系碩士班 === 94 === The southwestern plain of Taiwan is mainly covered by young Quaternary Holocence alluvium, composed of sand, silt, clay and gravel. As a soft ground in an engineering sense, highway embankments built on this region may experience severe settlement problems, so that settlement prediction plays an important role in the design stage. However, conventional settlement evaluation techniques are only accessible to consulting engineers but not transparent to field engineers. On the other hand, some literature studies have concluded that the difference between predicted settlement and in-situ measured one is usually considerable. The thesis is intended to establish a preliminary prediction model of embankment settlement primarily based on the data of boreholes, embankments, and field measurement along two east-west highways (Route 78: Taisi-Kukan and Route 82: Tungshi-Chiayi, abbreviated as HW78 and HW82, respectively), using the neural network analysis (NNA).
In this research, ten evaluation factors, among which six ground parameters require additional unification, were first selected from various ones, to predict the embankment settlement. Totally, 109 sets of data collected from HW78 were divided into three categories (I, II, III), each of which was adopted to perform the training part of NNA, followed by verification of the resultant model according to the remaining 8~16 data sets of HW78 and 8~13 data sets of HW82. The average and standard deviation of deviation percentage (m|e|,se) were defined for the checking purpose, and the results were compared with the consulting design and computation ones. The linear model of NNA was also used to determine the importance order of ten evaluation factors, from which the influence of reduced number of evaluation factors on the prediction model was studied with three modes (Modes 1,2,3 for 10, 8, 6 factors, respectively).
The verification results based on the remaining 8~16 data sets of HW78 depicted that (m|e|,se) of the NNA-generated nonlinear prediction model were both less than those of consulting design and computation ones (by around 50%), and it can serve as a quick tool in the near future for predicting embankment settlement in the studied area. The linear model of NNA pointed out the most important factor to be the fill height (h) for all three categories, implying the less variance of other factors in the studied area. On the other hand, in the verification stage using the 8~13 data sets of HW82, the ANN-generated nonlinear prediction model yielded a little higher (m|e|,se) than the consulting design and computation ones, indicating that the site condition along HW82 is somewhat different from that of HW78. It implies improving the applicability of this ANN model to the Chiayi area requires adding more data sets along HW82. In the study of reduced number of evaluation factors, reduction of the number of evaluation factors up to 4 (Mode 3) took a cost of increasing (m|e|,se) by about 30%; thus, it is up to the decision of site engineers to pick up of the right number of evaluation factors.
|