Rate effects on- and classification of penetration tests in Norwegian soils

Abstract

The aim of the PhD project has been to enhance knowledge on interpretation of commonly used soil investigation methods in Norway. It is divided into two parts, one where new soundings are performed to gain insight on the effects of deviating from standard penetration rates, and the other where large amounts of archived test results from field and laboratory are used to develop and tune new classification models. Penetration tests are commonly used in Norway to estimate soil stratification and, depending on equipment, measure the absolute or relative resistance between layers of the encountered soils. The most frequently used methods in Norway today are static tests, penetrating the soil at fixed rates. Some methods require an instrumented tip, while others measure resistance at terrain level. In practice, penetration rates vary due to equipment limitations, causing inaccuracies in measurements. To increase the value of archived tests, this project includes two studies aimed at better understanding the effects of small deviations from prescribed rates, which are commonly encountered in practice. The first study uses a simplified approach for the total sounding method, linking push force under different penetration rates to the expected push force using the standard rate. In all, 26 total soundings were conducted across two test sites. This approach is limited as total soundings only measure the penetration resistance at terrain level, which includes components from both tip- and rod frictional resistance. The analysis included manual alignment of resistance profiles to account for variations in layer thicknesses and inclinations between boreholes. The second study involved 63 cone penetration tests with pore pressure measurements (CPTu), conducted across three sites in central Norway. Considering observed penetration rates from over 2000 CPTu tests from road projects in Norway, a range of interesting penetration rates was identified. Tip resistance profiles were aligned to account for variation in layer thickness and inclination between boreholes. Errors observed in the source data for tests with faster than standard penetration rates were remedied by re-interpreting raw sensor data from the tests. Rate effects on tip resistance, sleeve frictional resistance and pore pressures were evaluated for the different soil types, as well as rate effects on soil behavior type classification using selected charts. Among algorithms tested to align tip resistance profiles in the second study was the dynamic time warping algorithm (DTW), which, in addition to curve alignments, produces a distance metric indicating the quality of the initial fit. This distance is used together with a machine learning classifier to define a soil behavior type (SBT) classification model for the total sounding method. Machine learning is especially useful for this task due to the abstract nature of the distance metric and the size of the dataset. With a large number of points divided unevenly across multiple soil classes, statistical methods were used to delineate zone boundaries of an SBT classification chart. The analysis showed that there is large overlap between point groups representing different soil classes. The classification study was repeated using different datasets to define models to screen for sensitive soils. A simple linear model was generated to identify sensitive soils with the total sounding, based on analysis of point density estimates. Similarly, point density estimates were used to define polyhedron boundaries in a three-dimensional soil behavior type classification model to screen for sensitive soils with the CPTu. The penetration rate studies show that rate deviations directly impact measurements, and that the degree of influence depends on both the type of soil and the magnitude of the deviation. If CPTu tests are performed within the prescribed tolerances of v = 20 +- 5 mm/s, rate effects on the tip and sleeve frictional resistance are expected to be within +-4%. Machine learning models were defined that produced superior soil behavior type classification results to those from the classification chart. However, these types of models require significantly longer evaluation times. The simple linear model for sensitivity screening produced results similar to those found with machine learning models. However, the model behavior can be controlled using a simple threshold parameter, making it attractive for practical application. Apart from machine learning models using the full dataset, classification models from this thesis have already been introduced into practice.