Decay‐like fracture diagnosis of composite insulator based on small sample virtual expansion and radar graph mapping

Abstract

Due to the complexity of test procedures and the high cost of measurements, the mechanism of the decay-like fracture is unclear, which brings hard concepts and the small sample problem. Therefore, it is intractable to establish accurate data-driven models. To solve the above mentioned problems, this paper proposes a data-driven modeling method based on radar graph mapping models and virtual sample generation (Radar-VSG). The main idea of the proposed method is to construct and select the most appropriate input features by using radar graph mapping, and to generate virtual samples by using the local outlier factor based on the isometric feature mapping algorithm (Isomap-LOF). To verify the effectiveness of our approach, a set of 27 training samples and 127 testing samples from China Southern Power Grid was applied. Experimental results suggested the accuracy was up to 100% after adding more than 27 virtual samples, with a 4.75% improvement. And the model has the best generalisation ability for 270 virtual samples, with the root mean square error of 0.018. Compared with support vector machine and four advanced VSG methods, the proposed Radar-VSG can achieve better performance.