Fault Detection of Hydroelectric Generators by Robust Random Cut Forest with Feature Selection Using Hilbert-Schmidt Independence Criterion
Yuki Hara, Yoshikazu Fukuyama, Kiyo Arai, Yuichi Shimasaki, Yuto Osada, Kenya Murakami, Tatsuya Iizaka, Tetsuro Matsui · 2021
This paper proposes a fault detection method for hydroelectric generators by robust random cut forest (RRCF) with feature selection using Hilbert-Schmidt Independence Criterion (HSIC). It is difficult to obtain fault data because faults unlikely occur in hydroelectric generators. Therefore, only normal data can be utilized for developing a fault detection method. Since data of hydroelectric generators have non-linear correlation, the fault detection method must consider the correlation. In this paper, the fault detection method of hydroelectric generators is assumed to be installed as cloud services in a data center. Therefore, there are two requirements for the applied method in order to develop a practical fault detection system. The first one is that the method has to generate an appropriate fault detection model for each hydroelectric generator in a short time in order to reduce CPU cost in the data center. The second one is that the limited number of features must be selected in order to save storage cost in the data center and communication cost through the internet. Therefore, if all data of hydroelectric generators are gathered, communication and CPU usage fees must be expensive. In order to reduce the costs, it is required to utilize only data with effective features for fault detection. The proposed method can construct models using only normal data, detect faults using data with non-linear correlation, and select effective features for fault detection of hydroelectric generators. The proposed method is compared with an isolation forest (IF) based method with all features, an IF based method with features of the highest accuracy, a RRCF based method with all features, and a RRCF based method with features of the highest accuracy. Area Under the Curve (AUC) values and calculation time are evaluated from the point of view of practical application of fault detection in data center. The proposed method is verified to be the most effective from the point of view. The results are verified using the Mann-Whitney U test.