Unsupervised Online Anomaly Detection on Multivariate Sensing Time Series Data for Smart Manufacturing

Abstract

The emergence of IoT and AI has brought revolutionary change in various application domains. One of them is Industry 4.0, also called Smart Manufacturing, which aims to achieve highly flexible and automated production processes. In this paper, we study a use case of anomaly detection in smart manufacturing using the real data collected from the sensing devices of a factory production line. Our goal is to improve the anomaly detection accuracy at an earlier stage of production line, so that cost and time wasted by possible production failures can be reduced. To overcome the limited and irregular anomaly patterns found from our multivariate sensor dataset, we proposed an unsupervised real-time anomaly detection algorithm based on LSTM-based Auto-Encoder. Our evaluations show that our approach achieved almost 90% accuracy for both precision and recall while other classification or regression based methods only reached 70%~85%.