Embedded autoencoder-based condition monitoring of rotating machinery

Abstract

Condition monitoring measures on industrial rotating machinery, such as pumps, turbines or generators are highly desirable to detect faults at an early stage and thus minimize operation-dependent downtimes, save costs and improve safety. In this article, an evaluation method implemented on an embedded system is presented that is able to detect the condition of rotating machines via vibration and acoustic measurement and to assess it using artificial neural networks. In terms of industrializability, an unsupervised learning method in the form of autoencoders is used, which is trained based on the nominal machine operation and embedded in a microsystem. Thus, the measurement system is primarily used for the identification of deviating machine behavior and threshold-based classification of operational capability. The system has been developed and validated using a test rig that simulates bearing damage and imbalances as defective conditions in addition to intact operation. This resulted in an F-score of 95.9 % of the applied smart condition monitoring system.