Effect of data drift on the performance of machine-learning models: Seismic damage prediction for aging bridges

Abstract

Machine-learning models play a crucial role in structural seismic risk assessment and facilitate decision-making by analyzing complex data patterns. However, the dynamic nature of real-world data introduces challenges, particularly data drift, which can significantly affect model performance. This adversely affects machine-learning models intended to aid emergency responders and disaster recovery teams. This study primarily focused on assessing the impact of column corrosion-induced data drift on the performance of machine-learning models for seismic risk assessment of bridges. The machine-learning model performance was evaluated with and without considering the impact of corrosion. The results revealed a significant decrease in prediction accuracy when the data drift effect was not considered. To address this challenge, this study proposes integrating principal component analysis-based anomaly detection to enhance the model performance. The optimized model considering drift demonstrates significant improvements in accuracy across corroded bridges aged 25, 50, and 75 years, with accuracy rates increasing from 90%, 85%, and 81% to 98%, 97%, and 96%, respectively.