Self-Supervised Dam Deformation Anomaly Detection Based on Temporal-Spatial Contrast Learning.

Area of Science:

• Geotechnical Engineering
• Structural Health Monitoring
• Machine Learning

Background:

• Dam deformation anomaly detection is vital for structural integrity and early warnings in dam health monitoring (DHM).
• Traditional methods require extensive labeled data, which is costly and labor-intensive.
• Unlabeled and semi-labeled data approaches are gaining traction due to data acquisition challenges.

Purpose of the Study:

• To introduce a novel spatiotemporal contrastive learning pretraining (STCLP) strategy for extracting discriminative features from unlabeled dam deformation data.
• To propose an effective anomaly detection method for dam deformation utilizing the developed STCLP strategy.
• To enhance downstream classification tasks through parameter transfer and prior knowledge fine-tuning.

Main Methods:

• Developed a spatiotemporal contrastive learning pretraining (STCLP) strategy combining spatial and temporal contrastive learning.
• Extracted discriminative features from unlabeled dam deformation datasets.
• Implemented an anomaly detection method by transferring STCLP pretrained parameters to downstream tasks and fine-tuning with prior knowledge.

Main Results:

• The proposed STCLP-based anomaly detection method demonstrated excellent performance in case studies involving an arch dam.
• The method significantly outperformed other benchmark models in dam deformation anomaly detection.
• Validated the effectiveness of extracting spatial and temporal features using the STCLP strategy.

Conclusions:

• The STCLP strategy provides a powerful approach for dam health monitoring using unlabeled data.
• The proposed method offers a robust and efficient solution for detecting anomalies in dam deformation.
• This research contributes to advancing data-driven techniques in structural health monitoring for critical infrastructure.