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

Area of Science:

• Civil Engineering
• Data Science
• Infrastructure Monitoring

Background:

• Structural Health Monitoring (SHM) is crucial for civil infrastructure safety and maintenance.
• Unsupervised anomaly detection is vital for identifying structural deviations without labeled damage data.
• Existing methods like Cumulative Distance Participation Factor (CDPF) and Semi-parametric Extreme Value Theory (SEVT) provide a baseline but struggle with subtle anomalies.

Purpose of the Study:

• To develop an improved unsupervised anomaly detection framework for SHM.
• To address the limitations of baseline methods in detecting subtle and non-linear structural deviations.
• To enhance the sensitivity, reliability, and interpretability of anomaly detection in civil infrastructure.

Main Methods:

• Implemented a baseline method combining CDPF and SEVT for thresholding using modal frequencies from the Z24 bridge dataset.
• Developed an ensemble anomaly detection framework integrating Principal Component Analysis (PCA) and Autoencoder (AE).
• PCA captures linear patterns, while AE learns non-linear representations for robust anomaly detection.

Main Results:

• The baseline method successfully identified anomalies in progressive damage scenarios.
• The proposed ensemble framework demonstrated improved sensitivity and reliability in detecting anomalies compared to the baseline.
• The ensemble method showed enhanced stability against environmental and operational variability.

Conclusions:

• Ensemble-based unsupervised methods offer significant advancements for SHM.
• The integrated PCA and AE framework provides a more robust and stable approach to detecting structural anomalies.
• This approach holds promise for more effective and reliable infrastructure health monitoring.