Physics-informed long short-term memory networks for response prediction of a wind-excited flexible structure. (15th January 2023)

• Record Type: Journal Article
• Authors: Tsai, Li-Wei; Alipour, Alice
• Is Part Of: Engineering structures. Volume 275(2023)Part A
• Publication Date: 2023-01-15
• Subjects: Flexible structures -- Wind-induced response -- Long-term monitoring -- Long short-term memory; Structural engineering -- Periodicals; Structural analysis (Engineering) -- Periodicals; Construction, Technique de la -- Périodiques; Génie parasismique -- Périodiques; Pression du vent -- Périodiques; Earthquake engineering; Structural engineering; Wind-pressure; Periodicals; 624.105
• DOI: 10.1016/j.engstruct.2022.114968
• Languages: English
• ISSNs: 0141-0296

Abstract

Highlights: A Long Short-term Memory (LSTM) architecture is used to predict the response of wind-excited structure using health monitoring data. LSTM that is trained with regular wind-excitations provide very good estimates of the response in extreme wind conditions. Limited sensor data was able to produce excellent estimates of the response in regular conditions. The approach can be used to predict response of other wind-excited structures under wind using limited data from sensors. Abstract: Slender and flexible infrastructures such as sign supports, cantilever traffic signal structures and high mast lighting towers are sensitive to wind force and were reported to have fatigue-related issues due to the large-amplitude vibrations throughout thier life. Simulating wind-induced structural response can be an important step to evaluate their fatigue life and reliability. However, wind simulations are usually quite complicated. A comprehensive wind force model was usually developed by conducting multiple wind tunnel tests. However, due to the high cost of wind tunnel tests and the limitation of a wind tunnel, aerodynamic and aeroelastic coefficients were usually extracted only at certain wind speeds and wind directions. Interpolation or extrapolation methods were commonly used when coefficients were not available, which makes the simulation result questionable. In this study, a methodology was proposed to simulate wind-induced structural response with lower costs. The proposed method uses monitoring data in the field to develop long short-term memory (LSTM) networks. In training LSTM networks, only the monitoring data in regular wind condition was used. However, the trained LSTM network can still predict the wind-induced response in high and extreme wind conditions observed during the monitoring of the structure. The proposed method can be useful when simulating wind-induced structural response in a wide range of wind speeds and can be widely used on other structures suspected of having fatigue damage due to wind-induced vibrations.