Anomaly Detection for Multivariate Time Series

Multivariate time series analysis provides early detection of anomalies in Oracle Customer assets and products, allowing customers to avoid costly failures and outages.

Project Overview

Our multivariate time series analysis is based on a nonlinear, nonparametric, multivariate pattern recognition technique called the Multivariate State Estimation Technique (MSET). It was originally developed to discover anomalies in time series sensor data in nuclear power applications and has evolved to big data prognostic applications commonly seen in safety-critical industries such as aerospace, utilities, and computer systems. MSET builds a model that predicts individual sensor readings based on the correlations among sensor data. It uses a statistical technique called the Sequential Probability Ratio Test (SPRT) to determine when the differences between predicted and actual values is significant. This provides the earliest possible anomaly detection and can even differentiate between sensor failures and monitored asset anomalies, providing potentially huge savings in maintenance costs.

MSET2 is an ecosystem that has been built around MSET. It provides a suite of data preprocessing algorithms to check for uncorrelated signals that can be monitored via univariate techniques, flat or ramp signals, data spikes (e.g., voltage) that need to be handled separately, and many other data input issues. It even has an algorithm to find and remove anomalies in training data. MSET2 also has a synthetic signal generator to obviate data privacy issues and a test harness that allows comparison of multivariate time series analysis techniques.