Method for detecting anomalies in multivariate time series data

摘要

A method detects anomalies in time series data, wherein the time series data is multivariate, by partitioning time series training data into partitions. A representation for each partition in each time window is determined to form a model of the time series training data, wherein the model includes representations of distributions of the time series training data. The representations obtained from partitions of time series test data are compared to the model to obtain anomaly scores.

主权项

1. A method for detecting anomalies in time series data, wherein the time series data is multivariate, comprising the steps of:
partitioning, based on time, time series training data into partitions, wherein each partition is treated as a separate and independent multivariate time series, and processing each partition using a sliding time window; determining a representation for each partition to form a model of the time series training data, wherein the model includes representations of distributions of the time series training data, wherein each distribution is a joint distribution over the time window of z(t), z(t+d) and an angle between the vectors (z(t), z(t+d)) and (z(t+d), z(t+2d)) for a single variable z(t) from the partition, where z is a variable of the partition and d is a delay; and comparing representations obtained from partitions of time series test data to the model to obtain anomaly scores, wherein the steps are performed in a processor.