Anomaly Detection in Aluminium Production with Unsupervised Machine Learning Classifiers

Abstract

This work presents a predictive maintenance methodology aiming at forecasting specific types of faults of an industrial equipment for anode production, utilizing process sensor data from operation periods. The challenge of this problem is the early detection of a fault, particularly just before it occurs. For the forecasting, some unsupervised machine learning architectures were tested. Several considerations were made for the pre-processing steps as well. Finally, automatic feature selection methods were introduced, one of which was used to find the most significant features within successive time windows of the evaluated historical data set. The experimental results, which conform to the visual observations, show that a warning time frame around 20 minutes before the incident is feasible for 43% of the incidents of a particular fault type within a critical 1.5-month period, whereas only in about 0.1% of the timestamps more than 75 minutes before such a fault an alarm is raised.