Anomaly Detection Features

(under construction)

SensorMINER makes it possible to automatically data mine archives and to generate a reasonable expert rule set for the purpose of real-time monitoring. SensorMiner is a time-series data-mining tool designed to handle data sets having specified characteristics. The goal was to build a human-readable real-time data-mining tool that can be used to help model and ultimately identify anomalous conditions that arise in real-world data mining domains. Using a combination of time warp, cluster, rule induction, Euclidian error, and state machine technologies, sensorMiner is a sophisticated temporal machine learner for real-time anomaly detection. 

Our view of time series anomaly detection is that of a machine learning or modeling task. Given a training set X of time series with an unknown probability distribution P, the task is to estimate P. Then given a new time series y, we assign an anomaly score inversely related to P(y).

With sensorMiner it is possible to construct working anomaly detection systems based on path or box modeling for this data set. Neither path or box models are appropriate for all time series. Some work may be required to tune parameters to a data set, but this is no different than most other anomaly detection systems. However these models have the desirable property that they can be visualized, which is an aid in verifying their correctness or modifying them manually to add domain specific knowledge.