Population-Based SHM Under Environmental Variability Using a Classifier for Unsupervised Damage Detection

Abstract

In this paper, we introduce a novel deep learning technique for anomaly detection in the context of Population-Based Structural Health Monitoring (PB-SHM). The proposed method eliminates manual feature engineering by utilizing Power Spectral Density (PSD) as input, allowing examination of the entire spectrum. It is based on an auxiliary classification task, wherein the model is trained to discriminate between different systems according to their dynamic response. The classifier confidence is then used during inference for damage detection. The neural network extracts discriminative features commonly impacted by damage, which are employed to create a normality model. The efficacy of our method is demonstrated on a simulated population of 20 individual 8- DOF systems influenced by a latent environmental variables, emphasizing its potential for PB-SHM under diverse conditions. Our technique achieves performance comparable to resonance frequency-based methods while potentially exhibiting higher capability in complex structures with multiple modes. Anomalies caused by a 5% decrease in stiffness are successfully detected, yielding an AUC of 0.94.