Speaker
Description
Prediction of dynamic behavior of multi-story timber buildings offers opportunities to reduce field measurements, extend structural lifespan by preventing cumulative damage, and enhance safety through real-time performance monitoring. While some environmental parameters such as moisture content (MC), relative humidity (RH), and temperature exhibit significant impacts on dynamic behavior of timber buildings, the use of insufficient and irrelevant data poses significant challenges for machine learning-aided regression modeling and data-driven prediction. When the major dynamic features of buildings are insufficient (e.g., modal frequencies from short-term monitoring programs) and the key measured environmental factors are weakly correlated with the dynamic responses, state-of-the-art regression techniques fail to capture real relationships between the environmental factors and dynamic responses and develop robust predictive models, leading to poor and unreliable predictions. To deal with these challenges, this paper proposes an innovative tree-based hybrid kernelized regressor called decision tree residual-embedded Gaussian process (DTRGP). This hybrid predictive method combines decision tree regression (DTR) and Gaussian process regression (GPR) within the realm of probabilistic kernel learning. The key purpose of this method is to develop a robust regression model with the highest prediction accuracy using insufficient and irrelevant data by leveraging the capabilities of some cutting-edge machine learning algorithms such as probabilistic kernel learning, residual learning, and hybrid learning. This method begins with initial regression modeling via the DTR by using available but inadequate and irrelevant data. Subsequently, virtual environmental data are provided by extracting residuals derived from the original and DTR-oriented predicted response data. Measured and virtual environmental data are combined to make enhanced information for developing the GPR and predicting the dynamic behavior of timber buildings. To validate the proposed method, this paper utilizes modal frequencies obtained from the Limnologen building, an eight-story cross-laminated timber structure located in Växjö, Sweden, along with environmental data recorded from a weather station. Results indicate that the proposed DTRGP method can considerably enhance prediction accuracy using limited modal frequencies and irrelevant information of recorded environmental factors.
