Realtime Modeling and Anomaly Detection in Multivariate Data Streams

This demonstration illustrates a realtime modeling and anomaly detection framework in cyber-physical systems using statistical models for multivariate data streams. Our framework uses a principled approach that utilizes a vector auto-regressive model fitted on a fast sampled data stream to learn a reduced order model of the system. Following that, we use Mahalanobis distances of the data stream residuals to detect anomalies using user-specified threshold. Various machine learning techniques are available to be selected by the user to classify the anomalies. Finally, the user has the ability to interface with the application by providing online reinforcement learning for the estimation and classification of anomalies in the streaming data.