Intelligent Cavitation Recognition of a Canned Motor Pump Based on a CEEMDAN-KPCA and PSO-SVM Method

• Published in: IEEE sensors journal Vol. 24; no. 4; pp. 5324 - 5334
• Main Authors: Li, Xiaojun, Yang, Haitao, Ge, Jie, Zhu, Shuai, Zhu, Zuchao
• Format: Journal Article
• Language: English
• Published: New York IEEE 15.02.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Adaptive sampling; Algorithms; Canned motor pump (CMP); Cavitation; cavitation recognition; Cooling systems; Data centers; Empirical analysis; Feature extraction; Heat transfer; Interference; kernel principal component analysis (KPCA); Liquid cooling; Overheating; Particle swarm optimization; Principal component analysis; Principal components analysis; Pumps; Redundant components; Sensors; support vector machine (SVM); Support vector machines; vibration; Vibrations
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2023.3347248

Due to interstitial flow and medium heat transfer, canned motor pumps (CMPs) are more susceptible to cavitation than other pumps. For CMP-driven liquid cooling systems, cavitation can reduce the heat transfer capability of the system or even lead to system overheating and failure. Therefore, it is necessary to monitor the cavitation status inside the CMP. In response to issues such as low computational efficiency and weak anti-interference capabilities, a new cavitation recognition method is proposed in this article. First, complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) combined with sample entropy is used to extract the cavitation features, and then, kernel principal component analysis (KPCA) is used to remove the redundant components in the features. Finally, the particle swarm optimization support vector machine (PSO-SVM) algorithm is applied to achieve the recognition of the cavitation states. Vibration signals in three directions of a high-speed CMP under different cavitation states are collected experimentally to construct a dataset for studying cavitation recognition. Experimental results show that the method proposed can effectively extract the cavitation features and reduce the redundant components, improving operational efficiency by 35.3% compared to the situation without reducing redundant features. The method also exhibits high recognition accuracy and better anti-interference ability. Under no interference conditions, it can achieve a recognition accuracy of 99.44%. Even under strong noise interference with an SNR of 6 dB, it can still maintain a recognition accuracy of over 90%.