Anomaly detection is a promising technique for monitoring the health of industrial machines, playing a crucial role in reducing maintenance costs and improving productivity. With the advent of online vibration acquisition systems, condition monitoring methods have evolved. The vast amount of data generated by these online monitoring systems necessitates automated anomaly detection methods to efficiently analyze vibration data with minimal human intervention. This article introduces a novel, fully automated, unsupervised approach for anomaly detection in machines using vibration data from autonomous wireless sensors. The method involves learning a reference model to analyze and evaluate deviations in new measurements. The approach is validated using real industrial vibration data. Results demonstrate its effectiveness in achieving a detection rate exceeding 90% while maintaining false alarms at less than 5%, thereby limiting the workload for the reliability team monitoring these machines. Furthermore, with expert feedback, the method can reduce the false alarm rate to almost zero and requires on average only 50 samples for training. Even if the first implementation relies on a client-server architecture, its low computational complexity makes it suitable for embedded implementations in resource-constrained environments.
Hachem et al. (Wed,) studied this question.