Enabling Damage Identification of Structures Using Time Series–Based Feature Extraction Algorithms

Abstract

Data-enabled approaches using statistical machine learning are emerging to help engineers identify damage among sensory data, which could be extremely helpful for decision making and data management in structural health monitoring. Although time-series data analysis using different feature extraction methods has been developed to improve data classification in machine learning, one of the remaining challenges is the spatial effects of sensory data. Methods that are suitable for individual sensor data are often limited due to their inability to account for system-level multivariate analysis of a group of sensory data. In this study, we attempted to develop a singular value decomposition (SVD)-based feature extraction method by designing a Hankel matrix to enhance multivariate analysis. We also presented a conventional autoregressive model (AR) and multivariate vector autoregressive model (VAR) for comparison to demonstrate the effectiveness of the SVD method for feature extraction. Further discussion was presented to qualitatively quantify the effects of uncertainty due to noise interference on the features and quantitatively evaluate the robustness of the feature extraction methods under different noise levels. The case study of a laboratory-based frame structure revealed that three feature extractions exhibited high capability in capturing designed damage scenarios, when sensory data was collected near the damage resources. However, the AR and VAR methods were both insensitive in detecting change when data were not near the event, which could be the case in real-world applications. By contrast, the SVD-based feature extraction method exhibited promising results for all cases.