An intelligent fault diagnosis method of hydroelectric generating unit based on enhanced cyclic spectral coherence and DCNN with group normalization

Abstract

Ensuring the secure and stable operation of hydropower plants relies heavily on effective fault diagnosis. In recent years, deep learning-based data-driven approaches have attracted significant attention for their strong ability to learn complex features. However, a key challenge lies in effectively integrating domain diagnostic knowledge into deep diagnostic networks to extract feature information closely related to the operating state of the unit and design more rational state predictors. To address this, the raw vibration signals are first converted into two-dimensional images using the enhanced cyclic spectral coherence (ECSCoh) method, while simultaneously achieving noise reduction and feature extraction, thereby reducing the difficulty of feature learning in deep diagnostic models. Secondly, the group normalization (GN) module is embedded into a deep convolutional neural network of Darknet-19 to normalize feature maps, mitigating internal covariate shift caused by data distribution discrepancies. Finally, by combining ECSCoh method with the GN-enhanced Darknet-19 network, a novel intelligent fault diagnosis model for hydroelectric generating units is proposed. Extensive testing on three datasets shows the model outperforms existing methods in accuracy and robustness, offering valuable technical support for practical fault diagnosis in hydropower plants.