Experimental Application of Self-organizing Feature Maps and Principal Component Analysis for Generator Condition Assessment

Data-driven approaches are gaining interest in the area of condition monitoring in electrical machines, because of the increasing availability of condition data - including the ability to capture such data - as well as the added flexibility over more traditional model-based approaches. Despite these benefits of data-driven approaches, the challenges of data imbalance, that is, the lack of availability of fault data as opposed to healthy operation data brings into question the suitability and practicability of deployment these methods. Thus, within the sphere of data-driven approaches, unsupervised learning or exploratory techniques could potentially lead to overcoming such challenges. This paper presents an experimental investigation into the use of self-organizing feature maps and principal component analysis for application in condition monitoring on a wound-rotor induction generator. A comparative analysis of these two techniques is conducted to determine the suitability thereof for condition assessment of the experimental generator under four different incipient fault cases. Both techniques are applied on experimentally measured voltage and current features. Results indicate that the self-organizing feature map technique do not yield suitable separation of condition clusters, whereas principal component analysis provides notably better performance.