A Stratigraphic Classification Estimation Method by the D-Layer Neural Networks

Information of borehole logs in a geotechnical information database has non-uniformity of spatial density of investigation locations, their depth and different appearance frequency of soils. So, it has characteristic of information bias. The purpose of this study is to propose a new idea of a global error function of Neural Networks (NNs) for stratigraphic classification estimation and its machine learning procedure in consideration with the information bias. Ordinary NN method (ONN) treats all the input data without discrimination in machine learning because the error function of NN is defined as a residual square sum in each data without taking account of information bias. In this study, Layer NN method (LNN) of which a new error function depends on an averaged residual squared sum in each stratum and distance of estimation location. LNN with and without consideration of dependence on the distance has been indicated as Distance Correlation Layer NN method (DLNN) and Simple Layer NN method (SLNN), respectively. In DLNN, the dependence on the distance expresses by Inverse Distance Weighting (IDW) method. Applicability and performance of ONN, SLNN, and DLNN in 3 cross sections in Fukuoka plain have been investigated by using a geotechnical information database. The proposed method that is DLNN has obtained higher correct estimation rate in each 3 cross sections than ONN and SLNN. And, it could eliminate or reduce incorrect estimations in the others. Therefore, the results have shown the proposed method (DLNN) is extremely effective for stratigraphic classification estimation by Neural Networks.