Mining multivariate time-series sensor data to discover behavior envelopes
Dennis DeCoste · 1997
This paper addresses large-scale regression tasks using a novel combination of greedy input selection and asymmetric cost. Our primary goal is learning envelope functions suitable for automated detection of anomalies in future sensor data. We argue that this new approach can be more effective than traditional techniques, such as static red-line limits, variance-based error bars, and general probability density estimation. Introduction 1 This paper explores the combination of a specific feature selection technique and an asymmetric regression cost function which appears promising for efficient, incremental data-mining of large multivariate data. Motivating this work is our primary target application of automated detection of novel behavior, such as spacecraft anomalies, based on expectations learned by data-mining large data bases of historic performance. In common practice, anomaly detection relies heavily on two approaches: limit-checking (checking sensed values against ty...