Improved ResNet network fusion transfer learning for hydraulic turbine state prediction

To predict the operational status of hydraulic turbines and mitigate the significant risks associated with turbine failures, this paper proposes an enhanced state prediction model based on an improved Deep Residual Network (ResNet) integrated with Transfer Learning (TL). First, the hydraulic turbine signal is decomposed using Variational Mode Decomposition (VMD), optimized by the Fruit Fly Optimization Algorithm (FOA) and reconstructed based on the kurtosis criterion. Next, the reconstructed signal is converted into a time-frequency representation using Continuous Wavelet Transform (CWT) followed by feature enhancement through histogram equalization. The ResNet architecture is further improved by incorporating skip connections, replacing standard convolutional modules with dilated convolutions to expand the receptive field. The traditional rectified linear unit (ReLU) activation function is replaced with a scaled exponential linear unit (SELU) activation function. Finally, the enhanced ResNet is combined with TL to establish a robust prediction model. The model's performance is evaluated using accuracy and recall metrics, demonstrating a high overall accuracy of 98.21 % and strong predictive capability for hydraulic turbine state monitoring.