Enhancing Condition Monitoring: A Semi-Automatic Framework Using Meta-Learning for Algorithm Selection

Condition-based maintenance systems play a major role in the enduring effort for modern production facilities to stay competitive. Driven by the ongoing digitalization of manufacturing systems, the aim is to increase the utilization of machines and their components and avoid unexpected failures during operation. Anomaly detection algorithms are the basis of a prognostics and health management process, but so far, the deployment of such algorithms is associated with significant labor from experts. Among other things, experts have to select appropriate data, find the best algorithms and their hyper-parameters, evaluate the performance and deploy a trained model for the online production environment. This paper presents a semiautomatic framework for the deployment of anomaly detection algorithms for manufacturing systems. It consists of a user interface that offers step-by-step guidance through the process of data-, algorithm- and parameter setup, as well as evaluation and export of a trained model. The core of the system is an advanced meta-learning-based algorithm recommendation method. Furthermore, the paper presents a user-oriented architecture that integrates all the necessary components, its aspects of IT security and shows general acceptance by a user experience test. The meta-learning recommendation system shows a significant reduction of the mean of the F-score deficiency towards the optimal selection of 25% compared to a benchmark selection.