Unsupervised Learning-Based Anomaly Detection for Bridge Structural Health Monitoring: Identifying Deviations from Normal Structural Behaviour

Abstract

Structural Health Monitoring (SHM) of large-scale civil infrastructure is essential to ensure safety, minimise maintenance costs, and support informed decision-making. Unsupervised anomaly detection has emerged as a powerful tool for identifying deviations in structural behaviour without requiring labelled damage data. The study initially reproduces and implements a state-of-the-art methodology that combines local density estimation through the Cumulative Distance Participation Factor (CDPF) with Semi-parametric Extreme Value Theory (SEVT) for thresholding, which serves as an essential baseline reference for establishing normal structural behaviour and for benchmarking the performance of the proposed anomaly detection framework. Using modal frequencies extracted via Stochastic Subspace Identification from the Z24 bridge dataset, the baseline method effectively identifies structural anomalies caused by progressive damage scenarios. However, its performance is constrained when dealing with subtle or non-linear deviations. To address this limitation, we introduce an innovative ensemble anomaly detection framework that integrates two complementary unsupervised methods: Principal Component Analysis (PCA) and Autoencoder (AE) are dimensionality reduction methods used for anomaly detection. PCA captures linear patterns using variance, while AE learns non-linear representations through data reconstruction. By leveraging the strengths of these techniques, the ensemble achieves improved sensitivity, reliability, and interpretability in anomaly detection. A comprehensive comparison with the baseline approach demonstrates that the proposed ensemble not only captures anomalies more reliably but also provides improved stability to environmental and operational variability. These findings highlight the potential of ensemble-based unsupervised methods for advancing SHM practices.

Keywords: SHM; Z24 dataset; adaptive weighting; anomaly detection; autoencoder; ensemble fusion; principal component analysis; unsupervised learning.

Figure 1

Comparative flow of the baseline CDPF–SEVT and the proposed PCA–AE Ensemble Fusion Framework.

Figure 2

The Z24 Bridge: (a) the longitudinal section, (b) the top view, (c) the pier and girder cross-view.

Figure 3

Dynamic response of the sensor in the Z24 bridge: (a) normal condition and (b) damaged condition.

Figure 4

ROC curve comparison between baseline CDPF-SEVT, AE-EVT, and PCA-AE.

Figure 5

Performance metrics comparison between baseline CDPF and ensemble fusion PCA-AE.

Figure 6

Anomaly detection based on threshold: (a) baseline CDPF-SEVT method, (b) ensemble fusion.

Figure 7

Stacked contribution area of PCA and AE for ensemble fusion.

Figure 8

Weighted score contribution: PCA vs. AE.

Figure 9

Empirical Cumulative Distribution Function of normalised anomaly scores for CDPF-SEVT and Ensemble Fusion methods.

Figure 10

False Alarm Rate comparison between CDPF-SEVT and Ensemble Fusion methods.