An LSTM-based anomaly detection model for the deformation of concrete dams

Abstract

Anomaly detection in deformation is important for structural health monitoring and safety evaluation of dams. In this paper, an anomaly detection model for the deformation of arch dams is presented. It combines the long short-term memory network (LSTM)-based behavior model for dam deformation prediction and the small probability method for control limits determination, and thus is called an LSTM-based anomaly detection model. To demonstrate the advantages of the LSTM-based anomaly detection model, the traditional hydrostatic-seasonal-time behavior model and the confidence interval method are considered for comparison. The 178 m-high Longyangxia Arch Dam is taken as a case study. The results show that the LSTM-based model has sufficiently high accuracy for dam deformation prediction, especially can accurately predict displacement peaks and troughs. The LSTM-based anomaly detection model can significantly avoid false warnings and missing alarms and is able to send alarms in time when the occurrence of adverse conditions causes abnormal deformation of the dam.