Interpretable damage sensitive feature extraction for drive-by structural health monitoring using knowledge distillation-based deep learning

Abstract

Drive-by Structural Health Monitoring (SHM) has the potential to effectively monitor the health conditions of bridges within transportation networks at a low cost. However, how to extract damage features from the drive-by measurements that are sensitive to changes in structural conditions while being robust to variations in vehicle loading scenarios and operational conditions is still an open question. This paper leverages advanced deep learning and knowledge distillation techniques to extract reliable damage-sensitive features from drive-by measurements in a supervised manner. To accomplish this, the proposed deep learning network is trained using thousands of drive-by measurements collected under various conditions, including different vehicle speeds, weights, directions, and structural damage states. This study includes an analysis of the network's intermediate layer outputs, numerical results from a finite element (FE) model of the test bridge, and theoretical derivations to interpret the physical significance of the learned damage features. The knowledge gained from this deep learning network subsequently informs and guides the theoretical development of generally applicable damage-sensitive features for drive-by SHM. In particular, this research work evidences that, in future drive-by SHM studies, more attention should be paid to the low-frequency driving speed-related component, instead of the conventional methods that focus on separating and extracting damage features from the bridge modal vibration related component of the drive-by measurement.