Digital signal processing technology is adopted to extract the characteristics of cavitation vibration signals of hydraulic turbines. Different statistical features of the signals in the time domain， frequency domain， and power spectral density are studied to detect the cavitation state of hydraulic turbines. The considered statistical metrics include: in the time domain， maximum value， minimum value， peak-to-peak value， mean value， variance， and mean square value； in the frequency domain， mean square frequency， centroid frequency， and frequency variance； as well as in the power spectral density， kurtosis， skewness， margin factor， and waveform factor. The feature signals include the first-order and second-order derivatives， which further enhance feature extraction. These features are input into the Extreme Gradient Boosting （XGBoost）， Recurrent Neural Network （RNN）， and Support Vector Machine （SVM） models respectively for classifying the cavitation states， and the accuracy of each model is calculated. The calculation results show that XGBoost achieves the highest accuracy， followed by RNN， while SVM has a relatively lower accuracy. Experimental results indicate that the identification of the cavitation state of hydraulic turbines using the proposed statistical metrics can reflect the cavitation stage of the turbine unit， verify the correctness of the proposed algorithm， and provide certain guiding significance for practical engineering.