TOWARD IMPROVED RISK ASSESSMENT OF INTERNAL EROSION-INDUCED EMBANKMENT FAILURE: A NEW NEURAL NETWORK APPROACH

Current internal erosion, risk assessment techniques for levees, tailings dams and other embankments tend to be limited and non-constitutive bringing their accuracy into question. Since Turnbull and Mansur’s (1961) pioneering work to characterize and assess failure risk for levees along the Mississippi valley 70 years ago, little progress has been made to adequately quantify risk beyond blanket theory and critical exit gradients, which have been demonstrated to be incomplete indicators of potential problems in certain cases. Clearly, present understanding of the role of significant factors in internal erosion including hydraulic gradients, seepage paths, insitu stresses, and soil mechanics is incomplete at best. This research presents a new data-driven modeling approach toward prediction of initiation of internal erosion using artificial neural networks in embankments. The focus is on initiation because both deterministic and probabilistic assessment tools rely on the factor of safety concept, and it makes sense to focus on preventing the initiation. Neural networks are an important computational technique whose functionality is based on the human nervous system. Major advantageous features of neural networks over both purely analytical models and purely statistical models include adaptation with presence of fresh data (i.e. “learn”), ability to implicitly detect nonlinear parametric relationships without prior assumptions, and ability to function when data are limited. Because the neural network technique is empirically derived, it provides both a systematic and quantitative basis for internal erosion initiation assessment without ignoring key underlying seepage theory and associated constitutive laws.