A deep residual shrinkage network based on multi-scale attention module for subsea Christmas tree valve leakage detection. (July 2022)

• Record Type: Journal Article
• Authors: Liu, Guijie; Chen, Yunqing; Zhang, Xiulong; Jiang, Weixiong; Xie, Yingchun; Tian, Xiaojie; Leng, Dingxin; Li, Zhixiong
• Is Part Of: Measurement. Volume 198(2022)
• Publication Date: 2022-07
• Subjects: Valve leakage -- Underwater acoustic -- Multi-scale attention module -- Deep residual shrinkage network; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2022.110970
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: A novel non-contact SCT leak detection method is designed. MSAM is designed to obtain rich sensitive features from strong background noise signals. An experimental platform for underwater valve leakage was built for the first time. Abstract: Recently, the valve leakage detection of subsea Christmas tree (SCT) attracts considerable attention in the field of underwater resource exploitation. However, most existing leakage detection methods rely on contact-sensors, which are expensive and troublesome to install and maintain. Additionally, it is still a great challenge to extract sensitive fault features for the non-linear and unsteady signal. To address these issues, a novel remote acoustic detection method based on acoustic sensor and deep learning is proposed in this paper. Firstly, the feasibility of SCT valve leakage detection based on acoustic sensors is theoretically demonstrated by acoustic analysis. Secondly, a multi-scale attention module (MSAM) is proposed to obtain rich feature information according to the characteristics of valve leakage acoustic signals. Subsequently, A deep residual shrinkage network based on multi-scale attention module (MSAM-DRSN) is designed to perform valve leakage detection. The proposed method is evaluated by the valve leakage experiment. The results show that the proposed leakage detection method can obtain sensitive fault features from strong background noise signals and the detection performance is better than existing detection methods.