This paper mainly reports on the training/generalization of experimental data on a scour-curve, which is a two dimensional locus remained when ice keel ploughs the seabed, by a Neural-Network (NN). As a result of training, the multiple correlation between estimated value by NN after training and experimental values was more than 0.99. We proposed that this method could replace a nonlinear multiple-regression analysis, which is very difficult to be applied when unknown parameters are independent. The NN model driven by an extensive data set will be useful tool for developing a practical method for estimation of scour depth by combining it with a mechanical ice scour model that we had already developed.
Ice-Scour Event is a phenomenon that occurs when ice comes into contact with seabed. Ice-Scour has been reported to cause damage to communication cables and water intake pipelines (Duval, 1975; Grass, 1984). While we have studied the scour-processes by many experiments on ice-scour event including the medium-scale model test, we have also developed the mechanical model[Kioka et al., 2000], which consists of the equation of motion concerning an ice and the simple model of interaction between the seabed and the ice keel. Also, we indirectly validated the mechanical model using the experimental results obtained in small- and medium-scale model tests [Kioka et al., 2001a; 2001b; 2002]. Our final goal is to develop the method for estimation of scour depth, or optimal depth of a buried structure such as a pipeline in order not to contact an ice keel. We believe that we can establish the method by combining the mechanical mod el that we had already developed and the scour curve. However, we need to establish the scour curve experimentally, which is difficult to be estimated by a theoretical approach.
