ANOMALY DETECTION OF SLOPE FAILURE BY ENSEMBLE LEARNING FOR PREVENTION OF OCCUPATIONAL ACCIDENT

Abstract

 Several measuring instruments and monitoring systems are being utilized to monitor slope deformations to prevent labor accidents due to slope failure. The issue for those instruments and systems is how to decide an “anomaly” using its monitoring data. Anomalies can be detected from the change in the strain of the slope surface using linear regression models and gradient boosting trees. This study attempted to improve the accuracy of anomaly detection using stacking method, a type of ensemble learning method. Centrifuge experiments of slope excavation were conducted, and anomaly detection was done using the data obtained. The slope was excavated for a total of 13 times and its deformation constantly monitored by eight strain sensors installed on the surface of the slope. In the experiment, slope failures occurred at the 6th, 8th, and 10th excavation cycles. The model was trained using the values before the slope failure, this established a state known as the normal state. Providing the model a threshold by defining a level or margin of difference, it can assess and validate whether a certain measured value was a normal or an abnormal value. By constantly monitoring the measured data and validating it with the predicted data, anomaly at any point in time can be recognized. Results from the analysis showed that performing stacking method improved the prediction accuracy compared to a single model and consequently lessened the number of false anomaly detections. This made the model more reliable to detect anomalies in slope failure.