Multiscale root mean square for hydropower generation systems: a scalable feature extraction method for condition-based maintenance

With the advancement of smart power stations, the maintenance paradigm for hydropower generation systems is shifting from scheduled routines to condition-based strategies, thereby posing new challenges for efficient anomaly detection from large-scale monitoring data. Existing signal feature extraction methods are largely based on nonlinear dynamic metrics, which identify anomalies in components like bearings and gearboxes by quantifying signal complexity. However, due to the abruptness and complexity of signals in hydropower systems, the effectiveness and applicability of complexity-driven methods in scenarios dominated by signal intensity variations remain to be fully validated. Therefore, the comparative assessment of intensity-based and complexity-based feature extraction approaches is essential for evaluating their suitability in extracting abnormal information from hydropower generation systems. Firstly, this paper summarizes various nonlinear dynamic and signal strength indicators, expanding them into multiscale indicators. Then, it compares their effectiveness in identifying abnormal processes in pump-turbine units. Subsequently, their performance in fault early warning scenarios is analyzed. Finally, the dynamic response of each method to different signal types under extreme load rejection conditions is observed. The results across various hydropower operation and maintenance scenarios demonstrate that the proposed signal intensity-based method named multiscale root mean square, significantly outperforms traditional entropy-based approaches, confirming its effectiveness and advantages. This study offers valuable guidance for the condition-based maintenance of hydropower generation systems and contributes to advancing the intelligence level of hydropower stations.