Anomaly detection in concrete dam using memory-augmented autoencoder and generative adversarial network (MemAE-GAN)

Abstract

Anomaly detection of concrete dam from deformation monitoring data is significant for dam safety evaluation. Existing anomaly detection models face challenges in identifying minor abnormal values and detection accuracy. This paper integrates the memory-augmented deep autoencoder (MemAE) with the generative adversarial network (GAN) to construct the unsupervised MemAE-GAN model, which leverages MemAE's precision in modeling and the GAN's adversarial training capability to highlight minor abnormal values, thereby significantly enhancing both sensitivity and accuracy in anomaly detection. Experimental results indicate that the MemAE-GAN model consistently achieved anomaly detection accuracy exceeding 0.97, considerably outperforming other comparative models. This model provides a highly accurate approach for deformation anomaly detection and lays the groundwork for subsequent research on deformation prediction and early warning. Future research could explore the algorithms to analyze the causes of abnormal values and establish the anomaly detection framework.