Cavitation State Recognition Method of Hydraulic Turbine Based on Energy-Entropy Features and Improved Stacked Denoising Auto Encoder

ObjectivesAiming at the problem that the acoustic emission (AE) signal induced by the cavitation of a hydraulic turbine is disturbed by background noise and the fault is difficult to identify, a state recognition method based on energy-entropy features and Harris hawks optimization (HHO) algorithm combined with 3-fold cross-validation (3Fold) to optimize stacked denoising auto encoder (SDAE) is proposed.MethodsFirst, the variational mode decomposition algorithm is used to decompose the signal, and then a series of intrinsic mode functions are obtained. Second, the energy and entropy features of the two intrinsic mode functions are extracted, which have the largest correlation coefficient with the original signal. A 12-dimensional feature vector is constructed and then the vector is imported into the recognition model. Third, the HHO algorithm combined with 3Fold is used to optimize the hyper-parameters of SDAE. Finally, the optimal parameters obtained by the HHO-3Fold-SDAE algorithm and other algorithms are input into the model for comparison.ResultsCompared with other algorithms, the HHO-3Fold-SDAE algorithm has smaller loss rate and accuracy variance, and higher average accuracy. Compared with SDAE, its test set average accuracy is increased by 6%. Compared with HHO-SDAE, its test set average accuracy is increased by 4% and accuracy variance is decreased by 17%.ConclusionsThe proposed method can be used to classify and recognize AE signals induced by cavitation of hydraulic turbine, and provide reference for the condition monitoring of hydraulic machinery.