Non-intrusive classification of gas-liquid flow regimes in an S-shaped pipeline riser using a Doppler ultrasonic sensor and deep neural networks

Abstract The problem of predicting the regime of a two-phase flow is considered. An approach is proposed that classifies the flow regime using Deep Neural Networks (DNNs) operating on features extracted from Doppler ultrasonic signals of the flow using the Fast Fourier Transform (FFT) is proposed. The features extracted are categorised into one of the four flow regime classes: the annular, churn, slug, and bubbly flow regimes. The scheme was tested on signals from an experimental facility. To increase the number of samples without losing key classification information, this paper proposes a Twin-window Feature Extraction (TFE) technique. To further distinguish the performance of the proposed approach, the classifier was compared to four conventional machine learning classifiers: namely, the AdaBoost classifier, bagging classifier, extra trees classifier, and decision tree classifier. Using the TFE features, the DNNs classifier achieved a higher recognition accuracy of 99.01% and greater robustness for the overfitting challenge, thereby showing the superiority of the DNNs in flow regime classification when compared to the four conventional machine-learning classifiers, which had classification accuracies of 55.35%, 86.21%, 82.41%, and 80.03%, respectively. This approach demonstrates the application of DNNs for flow regime classification in chemical and petroleum engineering fields, using a clamp-on Doppler ultrasonic sensor. This appears to be the first known successful attempt to identify gas-liquid flow regimes in an S-shaped riser using Continuous Wave Doppler Ultrasound (CWDU) and DNNs.