Boosting anomaly detection with unsupervised K-Means and SOM for energy-efficient factory machines

Radu Sorostinean, Catalina Neghina, Arpad Gellert

Abstract: This study presents a novel way to boost anomaly detection within industrial applications by using clustering algorithms and unsupervised approaches. The proposed hybrid technique relies on established methods in a novel pipeline composed of clustering and anomaly detection, resulting a unique way to monitor industrial production, improve sustainability, and prevent outages within Industry 4.0. The clustering component relies on an enhanced K-Means with automatic cluster detection and on a self-organizing map. The operating mode detection is designed to be used in a wider system, where smart factories can use live data from machines and assembly lines that are not designed with internet of things in mind to get preventive warnings about anomalies. After clustering, the anomalies can be identified in each operating mode more precisely. The approaches shown in this paper are evaluated through a case study from a woodworking factory that relies on energy and environmental sensors for different machines to detect in which mode the machine is working. The novel aspect presented is the use of unsupervised clustering methods to detect operating modes and boost classical anomaly detection on an assembly line via power consumption and environmental sensors. The results obtained show a boost in the performance of the classical anomaly detection algorithms. The clusters are closely correlated with the real operating modes of the machine, making thus easier to integrate older machines with the internet of things and smart factories. The anomaly detection applied on each individual cluster achieves significantly higher recall, reaching 1.0 in some configurations.

External IDs:doi:10.1007/s10845-025-02754-7