Research on a Centrifugal Pump Cavitation State Recognition Model Based on Autoencoder

The identification of cavitation state is one of the challenging aspects of centrifugal pump condition monitoring. In order to improve the effectiveness of cavitation state identification, a method based on autoencoder for centrifugal pump cavitation state identification is proposed. Experimental studies on cavitation of the centrifugal pump were conducted, and vibration signals of the pump casing at work frequencies were collected at three different temperatures. The original signals were subject to time-frequency domain analysis to obtain time-frequency graphs. The unsupervised training was used to learn the features of the input data, and an autoencoder was employed to identify the four types of cavitation states in the centrifugal pump. The research shows that this model can effectively extract the features of the centrifugal pump cavitation vibration signals and achieve a successful identification rate of 94.58%, which is 14.58% higher than the traditional Fast Fourier Transform-Support Vector Machine model. © 2024 Science Press. All rights reserved.