A robust approach for anomaly detection using 1D convolutional Siamese neural networks to enhance structural health monitoring in dams

Abstract

Accurate detection of structural anomalies in dams is critical for ensuring their safety, functionality, and longevity. This study introduces a novel approach utilizing a Siamese neural network (SNN) with 1D convolutional architectures to detect and classify anomalies in dams. The SNN leverages its unique capability to learn similarity measures between pairs of input sequences, enabling robust performance across both previously seen and unseen anomaly scenarios. The proposed model is rigorously evaluated against established machine learning methods, including Random Forest (RF) and Support Vector Machine (SVM). The results demonstrate that the SNN achieved superior accuracy across all unobserved classes, with values greater than 94.0%. In comparison, RF and SVM exhibited notable variability in performance, particularly for unseen conditions, underscoring the limitations of traditional approaches. By integrating 1D convolutional architectures, the SNN effectively captures temporal patterns and inter-sensor relationships specific to dam monitoring, enhancing its capability to detect subtle structural changes. These findings highlight the potential of deep learning-based approaches in advancing structural health monitoring practices for dams, paving the way for more reliable and precise anomaly detection systems tailored to the demands of critical infrastructure monitoring.