Fast Anomaly Detection For Multivariate Industrial Time Series Data

Abstract

Anomaly detection is used to identify/predict atypical patterns or outliers within datasets. It proves highly beneficial in numerous applications like industrial monitoring, healthcare and cybersecurity etc. With the advent of Industry 4.0, system automation is rapidly increasing, resulting in an exponential surge in data generation. In a factory setting, a sudden deviation in signal values can result in significant losses. Having a robust anomaly detection system in place becomes crucial to promptly identify and flag anomalies, thereby minimizing potential losses. In this paper, the proposal entails a down-sampling technique that offers advantages in terms of mitigating noise, reducing storage requirements, and diminishing computational demands by 88% for a representative dataset. Additionally, an anomaly interval detection algorithm designed to label data is introduced by us, thereby facilitating its application in supervised anomaly detection methods and predictive anomaly detection.