Vibration-based multiphase flow identification by deep learning for the vertical section of subsea pipelines

Flow pattern identification is critical for the flow assurance of the multiphase flow in the offshore oil & gas industry. For this purpose, an intelligent flow pattern identification model based on a convolutional neural network (CNN) is proposed in this study to identify different flow patterns of two-phase flow in the vertical section of subsea pipelines. The different vibration signals from four vibration sensors are converted by the continuous wavelet transform (CWT), and then fed into the improved LeNet networks, where the features in the last layer of the four LeNet are fused to develop the multi-input parallel convolutional neural network (CWT-MulLeNet). A series of two-phase flow pattern experiments for the vertical section of subsea pipelines are implemented in the multiphase flow loop to verify the performance of the proposed model. The results show that the accuracy of the proposed CWT-Mul-LeNet model is higher than that of CWT-LeNet (a single vibration sensor is allocated). Meanwhile, the performance of CWT is better than hilbert-huang transform (HHT) and short-time Fourier transform (STFT) in terms of time-frequency conversion. In addition, the identification accuracy of 99.06 % characterized by CWT-Mul-LeNet can be further improved by introducing the convolutional block attention module (CBAM) to 99.69 %, which is explained with the 3D t-SNE algorithm by means of feature visualization. The relevant data collected from the experiment can assist in the study of pipeline flow characteristics. The constructed model integrates information from complex positions, fully compensating for the shortcomings of traditional models with a single source of information data features, and improving the accuracy of intelligent flow pattern identification.