Prediction and detection of abnormal usage of an elevator

Abstract

In this thesis we used machine learning to detect the anomalous use of elevators by measuring the behavior at any specific time. We examined the data for unusual patterns in comparison with observed samples from the history of the elevator. We investigated forecasting the future use of the elevator to define an abnormal behavior for elevators and tried to address this issue in two approaches. First, we used Long Short-Term Memory to forecast future usage, and we optimized the result by extracting features and removing the noisy part of the data. Then we compared actual usage with our prediction and used 99.7% confidence interval (three sigma rule) to find out anomalies. The second approach used a local outlier factor to find out the distance of each week’s usage of the elevator from the other weeks. Then we took the intersection of these two methods, performed a set of post-processing actions to decrease the ratio of false positives and remove anomalies which were not sustained.