Unsupervised anomaly detection based on deep autoencoders, information fusion, and active sensing

Structural health monitoring plays a crucial role in ensuring the safety and resilience of engineering structures. Detecting structural anomalies is essential for maintaining the safety of citizens and the normal operation of civil infrastructures. In this study, a novel anomaly detection framework is proposed based on deep autoencoders (DAEs), information fusion, and active sensing. The framework involves exciting the structure at specific locations and collecting acceleration data. The data collected from multiple excitation and sensor locations are analyzed and fused to enhance anomaly detection performance. More specifically, an unsupervised anomaly detection framework using DAEs has been developed. Continuous wavelet transforms (CWTs) of acceleration signals are utilized to train DAEs. Information fusion strategies are proposed to enhance the robustness of the approach to both structural uncertainty and measurement noise. A comprehensive evaluation is performed to compare the performance of fully connected AEs, convolutional AEs (CAEs), variational AEs, transformer AEs, and one-class support vector machines, each trained separately on either CWTs or raw acceleration signals, to investigate the effect of input representation on anomaly detection performance. The results show that CAEs outperform other DAE-based approaches in detecting structural anomalies, achieving higher F1 scores and lower computational costs. Additionally, training DAEs with CWTs yields better performance than using acceleration time series. Numerical studies based on the ASCE benchmark structure and experimental studies based on a geodesic dome structure have been carried out to study the capabilities as well as limitations of the proposed approach. The effects of sensor and damage locations on anomaly detection performance are analyzed through damage identifiability. Solutions are proposed for the practical cases of limited sensors and insufficient data. The framework’s ability to extract information from multiple sources allows it to identify anomalies that traditional detection methods might have missed.