On-site Multi-Class Feature Selection for Online Classification of Switchgear Actuations in the Distribution Grid

Abstract:
The electrical grid is highly dependent on switchgear to maintain a safe and reliable power transmission. For this reason, the interest in on-site, non-invasive monitoring solutions including the detection of switch operations, their differentiation and ageing has significantly increased in the last years. Thereby, the research field of tracking acoustic emissions generated during the switching using low-cost micro-electro-mechanical system (MEMS) based sensors is emerging. This paper presents a computationally efficient method for selecting process- and design-specific features on-site (on a sensor system or gateway) to eliminate the need of prior offline training. This ensures generalized usability for different switch types and sensor positions without high re-training effort. The selected features are further used for online multi-class classification of switching processes. The proposed self-learning method, as well as the use of the MEMS sensors (acoustic and vibration), are both evaluated for classification performance on switchgear measurements during twelve different processes, leading to a robust classification with an accuracy of over 95 % in average.