An Adaptive Unsupervised Structural Health Monitoring Framework Using Exclusion-Aware Convolutional Autoencoders
TL;DR
This paper proposes an anomaly detection framework that combines the sequential retraining of CAE networks with an exclusion logic strategy, and demonstrates good generalizability, seamless adaptability to varying monitoring scenarios, and robustness across both training and validation datasets.
Abstract
Structural health monitoring (SHM) plays an important role in ensuring the safety and durability of civil structures. Among various SHM methodologies, autoencoders have proven particularly effective for anomaly detection due to their ability to extract meaningful features from complex datasets. This study focuses on convolutional autoencoders (CAEs) as an SHM tool, highlighting their efficiency in identifying structural changes, especially when fine-tuned with optimized parameters. The paper proposes an anomaly detection framework that combines the sequential retraining of CAE networks with an exclusion logic strategy. Two case studies are used to evaluate the approach: (i) a laboratory-tested aluminum plane frame subjected to five mass addition scenarios under impulsive loading; and (ii) the Z24 bridge, a full-scale structure exposed to simulated settlement damage. Key hyperparameters, including the latent space dimension, convolutional kernel size, and mini-batch size, were optimized to maximize model performance. Analyses were conducted in the frequency domain using the Mahalanobis distance as a statistical anomaly detection metric rather than as a measure of physical damage severity. Despite the deliberate exclusion of environmental variability, the proposed strategy demonstrates good generalizability, seamless adaptability to varying monitoring scenarios, and robustness across both training and validation datasets.