Baseline optimized autoencoder-based unsupervised anomaly detection in uncontrolled dynamic structural health monitoring

Abstract

Autoencoder reconstruction-based unsupervised anomaly detection is widely used in structural health monitoring. However, these methods typically require training on historical data from healthy structures, collected under environmental conditions similar to the test data. This limits their practical use, as it demands a comprehensive dataset of historical guided waves gathered across various environmental and operational conditions. Additionally, these methods fail when the training data contain a significant portion of damage-induced guided waves, as the autoencoder may reconstruct damaged waves just as effectively as normal ones. To overcome these challenges, our anomaly detection model is trained directly on current measurements, eliminating the risk of environmental discrepancies between training and test data. Furthermore, our baseline optimization strategy biases the autoencoder toward reconstructing normal guided waves, enabling reliable anomaly detection even when a large proportion of the training data are damage-induced waves. Additionally, we present a strategy to enhance the model’s practical performance by optimizing the weight factor for baseline loss and the baseline set size, based on guided wave reconstruction performance, without relying on damage labels. The effectiveness of this baseline-optimized autoencoder model, even when the training data contain significant damage-induced guided waves, is validated through measurements from 10 regions, each spanning 80 days of guided wave data collected under uncontrolled and dynamic environmental conditions.