Feature Selection Considering Characteristics of Operating Data and Random Cut Trees for Hydroelectric Generator Fault Detection

Abstract

This paper proposes feature selection (FS) considering characteristics of operating data and Random Cut Trees (RCTs) for hydroelectric generator (HG) fault detection (FD) by Robust Random Cut Forest (RRCF). If the faults are detected using knowledge of experts, huge efforts of the experts are required. Therefore, the faults should be detected using only data without knowledge of experts. The FD method has to treat non-linear correlation of HG operating data. Faults of HGs seldom occur and it is difficult to obtain fault operating data of HGs. Therefore, the faults must be detected by generating models with only normal data. In this paper, the developed HG FD method is assumed to be utilized in a data center for a HG FD service. Therefore, for reduction of communication and CPU costs, only effective features for HG FD should be selected. The proposed method can treat non-linear correlation of HG operating data. Without knowledge of experts, faults can be detected using only data. FD models can be generated with only normal operation data. Effective features for FD of HGs can be selected. The proposed method is compared with a RRCF based method using selected features by Hilbert-Schmidt Independence Criterion and k-nearest neighbor (HSIC-KNN-information coefficient and k-nearest neighbor (MIC-KNN-FS), and Isolation Forest (IF) and RRCF based methods using all features. The proposed method is confirmed to be more effective than other methods with calculation time and the Area Under the Curve (AUC) values. Moreover, the Kruskal-Wallis test and Mann-Whitney U test with Holm correction is applied to verify the results.