Fault Detection of Hydroelectric Generators using Isolation Forest

This paper proposes a fault detection method for hydroelectric generators using isolation forest. In hydroelectric generators, since faults rarely occur, it is difficult to obtain fault data. Therefore, it is required to construct models automatically using only normal data and perform fault detection. Hydroelectric generator data have non-linear correlation. Isolation forest can develop models automatically with only normal data and have been verified to be effective for the data with the characteristic. Namely, it is expected to be effective for fault detection of hydroelectric generators. Effectiveness of the proposed method is verified by comparing with a Multivariate statistical process control (MSPC) based fault detection method. Results indicate that the proposed isolation forest based fault detection method is more accurate than the MSPC based method. It is also verified using Wilcoxon signed-rank test.