Incipient Fault Detection of Power Distribution System based on Statistical Characteristics and Transformer Network

ncipient fault detection has become increasingly important in power distribution systems since higher requirements on power supply reliability were proposed. With more sensors and developments of intelligent algorithms, data-driven methods are recognized as a promising direction. In this paper, two main parts of this time series classification task are discussed: feature extraction of events and fault detection classifiers. For feature extraction module, statistical characteristics are comprehensively analyzed and chosen with hypothesis test and FDR multiple test. For fault detection classifiers, one of the most recent and powerful models, Transformer Network, is introduced with corresponding adjustments for the fault detection application. The inside attention mechanism and ability to build long-distance data dependency are the reasons for the high-level performance of Transformer. Experiments prove that the proposed framework is reliable in different conditions and has much better performance than related works, which provides an excellent reference for time-series classification tasks in the field of online monitoring and fault diagnosis.