Abstract: In the context of new power systems, wide-load operation of a hydro turbine has led to its increasingly complicated blade flow conditions and accelerated wear, making wear faults a major cause of its failure. From the acoustic perspective, this study develops a hydro turbine wear fault identification model based on the ensemble empirical mode decomposition with cumulative mean aggregation (EEMD) denoising, combined with a convolutional neural network (CNN) and a long-short-term memory network (LSTM). A raw signal is preprocessed by using EEMD to remove noise while retaining its key features. Then, we adaptively extract and reduce the dimensionality of the fault features by using CNN and input it into the LSTM model for feature learning and model training, achieving fault pattern identification. A hydro turbine fault test bench has been constructed for training and validation of the new EEMD-CNN-LSTM model. Experimental results demonstrate that this model is effective in wear fault identification, achieving an accuracy of 93.8% and outperforming the CNN, LSTM, or CNN-LSTM model. It improves accuracy by 6.7%, compared with the CNN-LSTM model without EEMD denoising. The findings are a valuable supplement to the previous studies of the monitoring and fault diagnosis systems of hydro turbines.

Key words: hydro turbine, denoising, deep learning, fault diagnosis, acoustic