Author(s): Yifan Yang
Linked Author(s): Yifan Yang
Keywords: Complex-pier scour; Data-driven methods; Deep learning; Hybrid neural networks
Abstract: This presentation will briefly introduce the history of research on the estimation of scour at complex bridge piers and highlight an emerging paradigm shift from using empirical methods to the utilisation of advanced data-driven, particularly deep-learning, methods. Estimating equilibrium scour depth and temporal scour evolution at complex piers have long been challenging due to the complexities in fluid-structure interaction, time-dependent structural exposure, grain size relative to large-scale turbulence, etc. Some recent works by the presenter and co-workers addressed the abovementioned complexities and proposed improved methods (Sheppard et al., 2023a, 2023b). Despite this advancement, the calculations are still affected by limited knowledge of physical processes and are not entirely compatible with modern automated systems. Recently, the presenter attempted to propose a hybrid neural network framework (Yang et al., 2024), including a multi-module multi-layer perceptron (MLP) network for predicting equilibrium scour depth and a long-short term memory (LSTM) network for temporal scour evolution forecasting, with extra data exchange between the two component networks for ensuring physical consistency. The multi-module MLP network passes 19 inputs into sub-networks representing different structural and flow factors for generating one output. The results show significantly better accuracy than existing empirical methods and standard simple neural networks. The LSTM network component provides multi-step-ahead scour evolution forecasting based on monitored time sequences combined with physics-related sequences from the interim and final outputs of the multi-module MLP model. Sensitivity analysis showed that an antecedent period of 48 h and a forecast horizon of 6 h may yield optimal model performance. For the capacity of conducting extended recursive forecasting, errors may accumulate after certain stages with scour-rate fluctuation given an inappropriate selection of antecedent period and forecast horizon. In general, the proposed data-driven model shows superior performance to traditional methods and good potential for integration with smart digital platforms. Finally, the presenter will make more discussions on the future directions in combining data-driven models and traditional empirical models towards more reliable and smarter pier-scour estimation and monitoring.
Year: 2024