RoSA:A Mechatronically Synthesized Dataset for Rotodynamic System Anomaly Detection

Abstract

The time-series datasets commonly applied for anomaly detection research showcase specific suboptimal properties. This work novelly conceptualizes condition state synthesis to improve the data-synthetic pipeline of an anomalous-event dataset. We demonstrate two technical contributions in this study. First, we propose a methodology to formulate, accelerate and enrich the condition state synthetic process. The proposed method includes three critical phases: analysis of a rotodynamic plant, systematic design of its condition state space, and development of a Markovian model for controlled state transitions. Second, a Rotodynamic System with Synthetic Anomaly dataset is constructed. It is a large-scale time-series dataset featuring controlled, abundant and diverse anomalous condition states, and per-time-step condition state labels. A comprehensive learning-based case study is conducted to illustrate that these unique features tangibly benefit anomaly detection research. Potential usages of the proposed dataset as an anomaly detection study benchmark are discussed.