Anomaly Detection for Strain of Slope Surface Using Machine Learning

As a soft measure to prevent sediment disasters, various measurement devices and monitoring systems have been developed through the improvement of ICT technology, and systems that detect signs of collapse and encourage evacuation have been developed based on slope observation data. A major challenge in those systems is what kind of measurement indicates a sign of danger. In this study, instead of predicting collapse based on a geotechnical model from measured data during slope failure, we modeled and analyzed observed data as time-series data to verify whether the signs before collapse can be captured as early as possible. As a method for evaluating slope observation data, we employed a learning method using slope observation data in a normal state during slope stability, and when a pattern of data different from the normal state was detected, we determined that the slope was unstable and issued an alert. A method for detecting anomalies on a slope was verified by using machine learning, in which data are predicted from a time series of slope surface strain data measured in a centrifuge field slope failure experiment, and by using the residuals between the predicted and measured data. As a result, it was confirmed from the time-series change in the number of anomalies detected by the eight installed surface strain sensors that anomalies were detected on the slope before the collapse.