Detecting Internal Erosion in Earth Dams: Frequency-Domain Analysis with Mel Spectrograms and Convolutional Autoencoders

Internal erosion poses a significant threat to the structural integrity of earth dams, often leading to catastrophic failures. The critical nature of such events emphasizes the need for reliable predictive methods to foresee and mitigate these risks. This study introduces a novel methodology for the early detection of internal erosion using deep learning (DL), incorporating a convolutional autoencoders (CAE) algorithm trained with Mel spectrograms of seismic monitoring signals. Mel spectrograms transform seismic time-series signals into a frequency-domain representation, which can improve DL model ability to detect subtle anomalies indicative of erosion. Incorporating data from the IJkdijk dam erosion experiment with 24 geophone sensors, this research develops an unsupervised anomaly detection model based on a multisensor analysis approach. This multisensor analysis aggregates data across all 24 sensors to holistically assess erosion initiation across the dam. To enhance interpretability, a confidence-based anomaly metric derived from Jensen–Shannon divergence is introduced, quantifying the progressive shift from normal to anomalous behavior. The initiation time and the first sensors to detect internal erosion are identified through this analysis. The outcome of the proposed frequency-domain-based approach matches the previous time-domain analysis conducted by the authors. Results demonstrate the consistency of the proposed CAE anomaly detection approach, capturing subtle yet critical early signs of internal erosion, which is crucial for timely intervention and mitigation strategies. This AI-based approach provides an innovative early detection/warning technique for monitoring earth dams, enabling timely intervention for internal erosion risk management.