Explainable Anomaly Detection in Industrial Streams

Anomaly detection in industrial environment is a complex task, which requires to consider multiple characteristics of the data from industrial sensors and anomalies itself. Such data is often highly imbalanced and the availability of labels is limited. The data is generated in streaming fashion, which means that it is unbounded and potentially infinite. The industrial process may evolve over time due to degradation of the asset, maintenance actions or modifications. The manual verification and definition of anomaly source is a tideous task, which requires human to carefully investigate each anomalous observation. An anomaly detection system should consider all above challanges. In this paper we propose a system, which addresses the discussed issues. It is applicable for industrial data stream scenarios and comprises of unsupervised anomaly detection model, resampling module and explanation module. We consider two different approaches towards the utilization of machine learning model – online and offline. We present our work in relation to a cold rolling process use case, which is one of the steps in production of steel strips.