Analysis of hydropower plant guide bearing vibrations by machine learning based identification of steady operations

Abstract

A novel machine learning based method is proposed to automatically identify steady operations of hydropower plants (HPPs) in this study. The approach applies the Pruned Exact Linear Time (PELT) algorithm to obtain the number of segments (steady operations & transients) for each working period by multiple change points detection in the HPP power output time series. An adaptive Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm, capable of self-adjusting its hyperparameters according to the PELT-defined segments, is then deployed for identification of steady operations. This adaptive characteristic can outperform other clustering methods in diverse HPP operational patterns through extensive comparison based on a three-year HPP measurement dataset and statistical tests. Based on the identification from the proposed method, the statistics of the HPP's upper guide bearing vibrations during both steady operations and transients before and after a known maintenance are compared, and an apparent bearing performance degradation can be revealed during signals from steady operations. It indicates that the proposed method can help to plan optimal bearing maintenance based on data of steady operations, and shows the potential for other practical applications for predictive maintenance of the different components of the HPP.