Prediction of Local Scour around Bridge Piers in the Cohesive Bed Using Support Vector Machines

Local scour around bridge piers is one of the most important factors threatening the life of bridges. The three-dimensional highly complicated horseshoe vortex and downflow are known to be the main agents responsible for pier scour. If the bed consists of cohesive sediment, it will add another level of complexity to the pier scour problem. Various approaches have attempted to predict scour depth, but no universal method is available to date. This study presents a prediction of local scour around bridge piers in the cohesive bed using support vector machines (SVMs), a machine learning technique. The maximum scour depth is predicted with seven dimensional variables, including velocity, flow depth, size of bed sediment, pier width, clay content, water content, and bed shear strength. The training and validation of the SVMs are conducted with 197 data from six datasets. Comparisons are made with the training and validation of the adaptive-network-based fuzzy inference system (ANFIS) method. The training of the ANFIS method appears successful, but the validation fails because of overfitting. The predictions with dimensionless variables are compared, and shown to be worse. In addition, the SVMs are found to predict the maximum scour depths better than three existing formulas, gene expression programming (GEP), and a non-linear regression model. The SVMs are applied to two datasets, revealing the importance of the coverage of the training data. Finally, to investigate the contributions of each variable, the mean absolute percent errors (MAPEs) and correlation coefficient are computed by predicting the maximum scour depths by excluding each variable.