Training sample balancing to train artificial neural networks for geotechnical problems

Abstract

This paper deals with the issue of balancing the training sample for training artificial neural networks when solving geotechnical problems. As an example problem, the prediction of mechanical characteristics of soil based on physical parameters is considered. A multi-layer full-connection artificial neural network was used to build the dependence. The need to improve the accuracy of analytical calculation methods for geotechnical problems is constant. Recently, such a regressor as an artificial neural network has been used more and more often for geotechnical problems. Neural networks are a powerful forecasting tool that allows reproducing dependencies of almost unlimited complexity. As neural networks need to be trained on a ready set of data, there is a question of quality of geotechnical test databases that are used for training. Due to the fact that there is no centralized way to collect geotechnical test data electronically, many researchers encounter significant incompleteness and imbalance in the data when attempting to collect such data. This paper proposes a solution for balancing such a training sample by generating examples of minority classes. It is proposed to balance the sample by generating missing examples with random parameter values within a certain range. The output data is proposed to be obtained with the help of existing calculation methods. This approach made it possible to make the training sample evenly distributed over the entire available range of values. At the same time, the range of predicted values increased in accordance with the limits of the experimental and generated sample. In addition, this approach allows us to take into account the existing analytical calculation methods when training neural networks.