One Class Support Vector Machine with a New Feature Selection Method for Fault Detection of Gas Turbine Generators in Thermal Power Plants

This paper proposes OCSVM with a new feature selection method for fault detection of GTGs in thermal power plants. There are three requirements for practical systems of fault detection of GTGs in thermal power plants. Firstly, it is necessary to detect fault using only normal data. Secondly, it is necessary to consider non-linear correlation of GTG data. Thirdly, limited features should be selected with only normal data considering development cost of the system. Using actual GTG data in thermal plant, effectiveness of the proposed method from the view point of calculation time and fault detection accuracy is verified. Especially for fault detection accuracy, OCSVM with the proposed feature selection method is compared with OCSVM with all features and OCSVM with the basic time-consuming feature selection method.