Unsupervised approach for online outlier detection in industrial process data
Michal Bechný, Johannes Himmelbauer · Procedia Computer Science · 2022
In this paper, we present a novel unsupervised approach for online outlier detection in multivariate streaming process data, which we developed in collaboration with our industrial company partner from the field of plastics industry. The main idea of our approach is to robustly capture the contemporaneous structure of the data and to use this to predict the future states, on the basis of which outliers are then detected. The prediction model is incrementally retrained on data from recent past, such that domain-specific outliers are identified, while meeting the requirements for evaluation speed. The main benefits of our approach are its applicability to heterogeneous data streams, real-time outlier detection, no need for historically labeled data, and its good performance for datasets with a high percentage of possibly consecutive outliers. In terms of prescriptive analytics, the algorithm can be used for real-time quality and process control. Furthermore, its graphical and numerical outputs can help to indicate by outliers most affected variables which can be related to a specific physical problem on the corresponding part of production machine.