Machine learning for bridge wind engineering

Abstract

Modeling and control are primary domains in bridge wind engineering. The natural wind field characteristics (e.g., non-stationary, non-uniform, spatial-temporal changing characteristics) and the wind-bridge interaction processes are physically complex and exhibit strong nonlinearity. The lack of analyzing these complex physical processes based on first principles makes it difficult for traditional modeling and control methods to accurately characterize or control wind field and wind-induced vibration of bridges. Data-driven automatic modeling and control is a new expansion direction in bridge wind engineering due to the extensive valuable historical data available and the complex, high-dimensional character of the state space and action space of the agent in active control. Machine Learning (ML) has been shown versatile for various modeling and control tasks in bridge wind engineering due to its excellent ability to automatically extract space features efficiently and make wonderful decisions facing complicated tasks without manual interval. Starting from an overview of the basic concepts and methods of ML, this review will comprehensively analyze and summarize the application of ML methods in various aspects of bridge wind engineering, including wind fields, static aerodynamic reconstruction, wind-induced vibrations and control based on reinforcement learning, in order to enable readers to have a full knowledge of the potential of ML in various subfields. The essence of these tasks is to rely on data-driven minds to solve tricky problems that cannot be solved based on first principles. ML will continue to integrate and coexist with wind engineering of bridges and generate new revolutions in both theories and applications owing to its powerful information-processing and modeling-learning ability.