ABSTRACT
Cone Penetration Tests are widely used in the Netherlands, due to their ease of execution in the Dutch delta deposits and their relatively low cost. As the amount of performed CPTs increases, an automated soil interpretation becomes more and more relevant. Attempts to automate soil classification have been done in the past, but the empirical formulas commonly used do not always provide a satisfactory interpretation for engineering purposes. Besides that the soil type is often not interpreted correctly, there is also the problem that the classification is provided for each measurement (every 2 centimeters) and no strategy is provided to aggregate those tiny layers. This paper shows how a data driven approach can yield better results than the traditional empirical methods. A machine learning model is presented which is based on a Convolutional Neural Network. The Neural Network has been trained on 1800 pairs of CPTs and boreholes that met the condition of being less than 6 meters apart. An algorithm based on the theory of signals is applied to the classification given by the model to group measurements into soil layers. The paper explains the theory behind the model, shows a comparison with the soil classification given by the Robertson correlation and shows how the model can be used in the geotechnical practice.
