Acoustic Emission-Based Measurement and Weak Fault Identification of Hydro-Turbine Runners Using a CNN-GRU Framework

Abstract

Early identification of weak faults in hydro-turbine runners is essential for safe and stable operation under complex hydraulic conditions. Conventional monitoring signals are often insufficient for detecting incipient faults such as runner cracks due to their weak and easily masked characteristics. In this study, an acoustic emission (AE) signal measurement approach is adopted, and mechanism analysis combined with operational experiments is conducted to investigate the acoustic characteristics of cracked runners. Cracked runners generate stable mid- to high-frequency AE features, with dominant energy concentrated in the range of 200–300 kHz, associated with transient elastic wave emissions induced by microcrack propagation and local stress concentration. Under both clear-water and sediment-laden conditions, cracked runners exhibit higher transient energy release rates and more pronounced pulse aggregation characteristics than normal conditions. A convolutional neural network (CNN)-gated recurrent unit (GRU) model with a self-attention (SA) mechanism is developed for weak fault identification. The proposed model outperforms conventional methods in accuracy and stability. The model achieves accuracies of 91.6% and 92.3% on the training and test sets, respectively, with fast convergence and strong robustness. This study provides an AE-based measurement and identification framework for early fault detection in hydro-turbine runners.