MADFlow : Normalizing Flow를 활용한 다변량 시계열 이상 탐지

초록

With the recent advancement of smart factories in manufacturing processes, high-dimensional data is being collected in real-time from multiple sensors in production facilities. However, it is very difficult to detect anomalies that reflect both correlations and temporal dependency between high-dimensional variables. In this study, we propose Multivariate Time Series Anomaly Detection via Normalizing Flow (MADFlow), which can reflect both correlation between variables and temporal dependency. MADFlow consists of a temporal encoder to reflect temporal dependency and a flow module to learn the distribution of high-dimensional data and is trained in an end-to-end manner. Experimental results on multivariate time series data with similar characteristics to data generated in manufacturing processes show that MADFlow has significantly better anomaly detection performance than existing models. Therefore, we expect MADFlow to be able to efficiently detect anomalies in real-world manufacturing processes.

키워드