Cavitation state recognition method of centrifugal pump based on multi-dimensional feature fusion and convolutional gate recurrent unit

Abstract

The rapid and accurate recognition of cavitation in centrifugal pumps has become essential for improving production efficiency and ensuring machinery longevity. To address the limitations of existing methods in terms of applicability, accuracy, and efficiency, a new method based on multi-dimensional feature fusion and convolutional gate recurrent unit (MCGN) was proposed. Experimental monitoring of cavitation of centrifugal pumps was conducted. Five signals at different water temperatures and operating frequencies were collected. Key modulating features were extracted by time-frequency analysis and principal component analysis. The multi-dimensional features are fused by one and two dimensional convolutional neural networks. The cavitation state label was used to label the sample set by cavitation number, net positive suction head, and cavitation evolution images captured by high-speed cameras. Finally, the neural network based on the convolutional gate recurrent unit was used to classify the cavitation state of the centrifugal pump. The experimental results demonstrate that the proposed method achieves recognition accuracies exceeding 98% for vibration signals, noise signals, outlet pressure pulsation signals, and torque signals. Compared with the short-time Fourier transform-autoencoder model, MCGN model can improve the recognition accuracy by 4.03%, computation efficiency by 20%, and loss by 87%. These advances underscore the potential of the method in monitoring and maintenance practices for centrifugal pumps.