Spatiotemporal Anomaly Detection and Prediction of Anomaly Propagation Path Using LSTM Networks

Abstract

Anomaly detection for connected systems is challenging as it is hard to observe and analyze multiple spatial and temporal scales of sub-processes and operations given the heterogeneity and sub-system level variances involved. For such dynamically evolving, non-stationary systems there are ongoing uncertainties involved in inherent data distributions. This thesis looks into characterizing spatial influences that govern large-scale system dynamics by exploiting sub-system level correlations capturing neighborhood specific information. The proposed framework is tested upon the entire traffic network of Nashville, TN focusing on developing mechanisms to identify sources of anomalous traffic behavior leading to congestion. In addition to this, a data-driven approach to capture the dynamics of traffic congestion propagation is deployed by creating a citywide ensemble of connected Long Short Term Memory (LSTM) network models that takes into account road-intersection specific information.