Abstract: Most of the vibration monitoring signals of hydropower units contain a large amount of abnormal data, which severely affect the assessment and prediction of unit health status. Considering the relationship between the unit vibration and its working condition, this paper presents a multi-scale cleaning method of unit vibration signals based on Gaussian Mixture Model and Density Based Spatial Clustering Applications with Noise (GMM-DBSCAN). First, DBSCAN is used to clean initially the vibration anomalies in the whole range of working conditions, and GMM is used to calculate the probability of different working conditions and divide their intervals. Then, for each of the intervals, a density clustering cleaning threshold for DBSCAN is calculated using its probability density, and its abnormal vibration data are cleaned. This method has a cleaning rate of abnormal data of up to 6.3‰, which has been verified using the vibration data monitored at the Pubugou hydropower station under the different working conditions of its one-year operation. Meanwhile, artificial anomaly data are used to verify the method. The results show the method can effectively clean out the isolated outliers and dense abnormal points of unit vibration signals, thus improving the health status evaluation and prediction of hydropower units.

Key words: hydropower units, vibration, work condition classification, unsupervised clustering, data cleaning