Artificial neural networks for the interpretation of piezometric levels at the rock-concrete interface of arch dams

Abstract

The aperture of the rock-concrete interface of arch dams can be characterised by analysing local piezometric levels. This highly non linear phenomenon involves thresholds effects. Consequently, it is poorly described by additive models such as HST (Hydrostatic, Season, Time), which is the classical multi linear regression model used in dam monitoring to assess the safety of structures. The presented study applies a more flexible statistical method, the Artificial Neural Network (ANN), so as to interpret uplift pressures in a French arch dam, via piezometric measurements. A thorough sensitivity analysis (SA) is performed in order to diagnose the evolution of the aperture of the interface. The originality of this work lies in the proposed methodology used to perform this SA avoiding extrapolation as much as possible, and on the tuning of the network which is implemented with a parametric study that integrates physically interpretable elements, as a supplement to the classical quantitative metrics. ANN turns out to be highly efficient and interpretable when used to study non linear phenomena. Finally, the gain that ANN brings to operational monitoring is discussed.