Cavitation Identification Method of Centrifugal Pumps Based on Signal Demodulation and EfficientNet

The recognition of cavitation status is crucial in the state monitoring of centrifugal pumps. For improving the efficiency of identifying cavitation status in centrifugal pumps, a unique method is proposed based on signal demodulation and efficient neural network (DEN). Experimental investigations of cavitation phenomena were conducted on centrifugal pumps. Vibration signals at six distinct frequencies were collected from the pump casing under three different temperature conditions. Signal demodulation was used to extract the characteristic frequencies of the modulated components. The preprocessed data were then input into a deep learning model that integrates MBConv architecture. Subsequently, the researchers conducted parameter optimization and cross-validation to develop the final DEN cavitation status identification model. The research results indicate that this method achieved a successful cavitation state identification rate of 89.44%. Compared to using FFT-transformed frequency domain signals as model inputs, the recognition accuracy improved by 20.69%. Compared to an autoencoder model, the recognition precision enhanced by 25.28%. The results confirm the efficacy of integrating signal demodulation and deep learning for cavitation recognition, providing a new technological pathway for the monitoring of centrifugal pump conditions.