YASA: Yet Another Time Series Segmentation Algorithm for Anomaly Detection in Big Data Problems

摘要

Time series patterns analysis had recently attracted the attention of the research community for real-world applications. Petroleum industry is one of the application contexts where these problems are present, for instance for anomaly detection. Offshore petroleum platforms rely on heavy turbomachines for its extraction, pumping and generation operations. Frequently, these machines are intensively monitored by hundreds of sensors each, which send measurements with a high frequency to a concentration hub. Handling these data calls for a holistic approach, as sensor data is frequently noisy, unreliable, inconsistent with a priori problem axioms, and of a massive amount. For the anomalies detection problems in turbomachinery, it is essential to segment the dataset available in order to automatically discover the operational regime of the machine in the recent past. In this paper we propose a novel time series segmentation algorithm adaptable to big data problems and that is capable of handling the high volume of data involved in problem contexts. As part of the paper we describe our proposal, analyzing its computational complexity. We also perform empirical studies comparing our algorithm with similar approaches when applied to benchmark problems and a real-life application related to oil platform turbomachinery anomaly detection.