Cavitation state recognition for control valve using AlexNet-type neural networks with three-channel images transformed by time series. (1st May 2023)

• Record Type: Journal Article
• Authors: Wei, Jianqiu; Liu, Xiumei; Li, Beibei; Zhang, Yujia; Shang, Ximing
• Is Part Of: Measurement science & technology. Volume 34:Number 5(2023)
• Publication Date: 2023-05-01
• Subjects: cavitation state recognition -- CNN -- multiple measuring points -- data fusion; Physical measurements -- Periodicals; Scientific apparatus and instruments -- Periodicals; Equipment and Supplies -- Periodicals; Science -- instrumentation -- Periodicals; Technology -- instrumentation -- Periodicals; Mesures physiques -- Périodiques; Physical measurements; Scientific apparatus and instruments; Periodicals; 502.87
• DOI: 10.1088/1361-6501/acb458
• Languages: English
• ISSNs: 0957-0233

Abstract

Abstract: The control valve is a crucial component with high switching frequency in coal liquefaction systems. Its performance exerts a great influence on the long-term operations in the coal chemical technology. Cavitation is a key factor affecting the control accuracy and surface morphology of the control valve; hence, it will further shorten the service life of the control valve. A cavitation state recognition model based on vibration acceleration time series is proposed in this paper, and the cavitation stage in the valve is identified accurately. Three vibration acceleration sensors are attached to the valve seat to measure cavitation-induced vibration data. The collected time series is divided into fragments based on the non-overlap moving window method, and these fragments are converted into feature maps using the Gramian angular field transform. These feature maps are then input into AlexNet to train the cavitation state recognition model. The experimental results show that recognition accuracy could be improved effectively upon the increase in the length of the time series fragment, and the proposed model has a similar prediction accuracy on unfamiliar datasets. Compared with no noise, the recognition accuracy could reach to 95% when the moving window length is 150 and the signal-to-noise ratio is equal to 5 dB. Furthermore, the proposed model could still achieve good recognition results under mixed open conditions.