ANOMALY DETECTION IN SLOPE SURFACE STRAIN USING DEEP LEARNING

Abstract

The operation of a monitoring system based on the detection of signs of collapse from slope observation data is effective as a soft measure to prevent landslide disasters. The major challenge of this monitoring system is to define how to trigger an alarm for evacuation when the measured data changes. In this study, we use time series data of slope surface strain measured in a slope failure experiment at the centrifuge modeling. We used LSTM, one of the deep learning methods, to predict the data, and verified the method to detect the anomaly of the slope by the residual between the predicted and measured values. From the time series of the number of anomalies detected by the eight sensors, it was confirmed that the anomalies could be detected before the slope collapse. In addition, the surface strains were converted to velocities in order to make the time series data stationary. In this case, it was confirmed that the anomaly of the slope could be detected before the collapse.