Fault identification of hydropower unit based on time–frequency feature map of vibration signals and Convrbm-ResNet

Abstract To address the significant impact of the non-stationarity and nonlinearity of vibration signals on the accuracy of fault identification in hydropower units, a method for condition identification based on time–frequency characteristics of vibration signals and Convolutional Restricted Boltzmann Machines-Residual Network (ConvRBM-ResNet ) is proposed. The vibration signals are first decomposed by adopting Complementary Comprehensive Empirical Modal Decomposition to further obtain Intrinsic Mode Functions (IMFs) at various frequencies. Then, by combining correlation analysis, the more sensitive IMF components are extracted. Furthermore, the obtained effective IMF components are converted into time–frequency feature maps, which are consequently adopted as inputs to pre-train the ConvRBM model to simplify feature representation. Finally, using the fault category as the output, a fault identification model for hydropower units is obtained by training ConvRBM-ResNet. The proposed fault identification model is validated with actual operation data of a hydropower station. The results indicate that the proposed fault identification model can accurately identify the operating status of the hydropower unit, achieving an accuracy of 98.88%, which is a 4% improvement over the non-improved method. The time cost is reasonable, and the model demonstrates strong robustness.