Active Learning for Condition-Based Maintenance of Industrial Machinery Using COMETH

We present a system for condition monitoring of industrial processes and machinery, called COMETH, which includes an active learning approach. The system is based on two complementary machine learning methods which are continuously updated through a feedback provided by the user of the system. The combination of two methods with active learning allows for the monitoring system to quickly adapt to changes in the boundary conditions of the application and to detect novel anomalies while keeping the amount of required feedback low. We illustrate the implementation and the usage of this active learning procedure for a conditioned-based maintenance of a towel picking machine. Additionally, we demonstrate the ability of the approach to correctly identify a detected anomaly. We show that the practicability for employing the proposed condition monitoring system can be further increased by using pre-trained models on similar machines in order to reduce the amount of required feedback and by providing a graphical user interface to facilitate the interaction with the technician.