Interpretation of P/CPT data using data fusion techniques.

摘要

Although in situ tests, such as the cone and piezocone penetration tests (P/CPT), have several advantages over traditional methods of sampling and laboratory analysis in determining representative properties of a soil deposit, the existing methods used to infer soil properties from P/CPT data are not always reliable due to the complexity of cone penetration. It is proposed herein that the process of data fusion can be used to estimate soil properties such as composition, overconsolidation ratio (OCR), coefficient of lateral earth pressure at rest (Ko), and undrained shear strength (su), directly from in situ test measurements, and that data fusion algorithms, through training, may be able to overcome some of the limitations of the current P/CPT interpretation methods.; To demonstrate that data fusion can be a useful tool for estimating soil properties from P/CPT data, databases consisting of P/CPT measurements and corresponding (known) values of various soil properties as determined in the laboratory were used to train and test several different data fusion algorithms, including the general regression neural network (GRNN), regression trees, and model trees. Two additional data fusion techniques, namely bootstrap aggregation and stacked generalization, were employed in an attempt to improve data fusion model performance. Additional features were created from the original set of P/CPT data based on the work of previous researchers in another attempt to improve the predictive reliability of certain data fusion models. Specifically, measured values of cone resistance and sleeve friction obtained from cone penetration test (CPT) soundings, together with grain-size distribution results of soil samples retrieved from adjacent boreholes, were used to develop a GRNN-based data fusion model for predicting soil composition from CPT measurements. Corrected cone resistance and pore pressures measurements obtained from piezocone penetration test (PCPT) soundings, together with one-dimensional consolidation and triaxial compression test results, field vane shear test results, and empirically-estimated values of Ko, were also used to develop both GRNN-based and tree-based data fusion models for predicting OCR, su and Ko from PCPT measurements. To demonstrate the benefit of fusing multisensor data, data fusion models were often developed using various combinations of P/CPT measurements and model performance was evaluated. Data fusion model predictions of soil properties were compared with the estimates obtained using existing interpretation methods to determine if the reliability of inferred soil properties can be improved by using data fusion techniques.; Through these analyses, data fusion was found to be an effective method for inferring soil properties from P/CPT measurements. The profiles of soil composition estimated by the data fusion model were found to compare generally very well with the actual grain-size distribution profiles and the results of two existing CPT soil classification methods; and the values of OCR, K o, and su predicted by the data fusion models were found to compare very well with the reference values, and to be generally more reliable than the results of the corresponding interpretation methods (those using the same PCPT data inputs). Fusing the features extracted from data obtained using two or more piezocone sensors tended to improve the reliability of the soil property predictions, and the use of additional created features often further improved soil property predictions. Thus, data fusion techniques may represent an improvement over the methods currently being employed to interpret piezocone penetrometer sensor data. Because the data fusion algorithms have the ability to deal with noisy training data, they can be very effective in modeling nonlinear multivariate problems and may be able to "learn" some of the complex nonlinear relationships (such as soil fabric, sensitivity, mineralogy, aging, etc.) among sampl