VMD-SE-DCNN-CBAM: An Intelligent Cavitation Recognition Method for Canned Motor Pump

Canned motor pumps (CMPs) are extensively used in the liquid cooling systems of high-performance computing (HPC) systems due to their leak-proof design, low noise, compact structure, and high reliability. However, CMPs are more susceptible to cavitation compared to other pump types, primarily because of their structural integration and greater gap flow. Cavitation can limit pump delivery capacity, reduce cooling efficiency, and lead to system overheating, making it crucial to monitor the internal cavitation state of CMPs. In this work, we propose a novel cavitation recognition method that combines variational mode decomposition (VMD) and sample entropy (SE) with a deep convolutional neural network (DCNN) enhanced by convolutional block attention modules (CBAMs), referred to as VMD-SE-DCNN-CBAM in this article. Specifically, intrinsic mode functions (IMFs) are efficiently extracted from vibration signals using VMD combined with SE. These IMFs serve as inputs to a DCNN integrated with CBAM, which is then used for cavitation recognition. A laboratory-scale cavitation recognition experimental setup is established to demonstrate the capability and effectiveness of the proposed VMD-SE-DCNN-CBAM method. Results show that the VMD-SE-DCNN-CBAM method achieves a remarkable recognition accuracy of 99.66% under four different cavitation states, outperforming other comparative methods.