Passive flow regime identification based on synchrosqueezing wavelet transform and deep learning

Two-phase flow produces flow patterns inside a straight horizontal pipe based on inlet velocities. Conventionally flow regimes are identified by comparing superficial velocities with available flow regime maps. A non-invasive technique of flow regime identification is required for the two-phase flow industry. This research proposes a passive method for automatically identifying flow regimes by combining synchrosqueezing wavelet transform (SWT) and deep learning. This study introduces a method for categorizing pipe surface vibrational displacement signals to identify flow regimes. The approach utilizes post-signal processing techniques, specifically continuous wavelet transform (CWT) and synchrosqueezing wavelet transform (SWT). It establishes a bijection relationship between flow regimes and the time–frequency (TF) textures extracted from passive displacement signals. Initially, the displacement signal of the flow regimes is captured using an accelerometer, followed by conversion into the time–frequency domain (TFD) via SWT. Subsequently, three convolutional neural network (CNN) architectures, namely AlexNet, VggNet, and ResNet, are trained to automate flow regime identification. The study investigates the impact of TF analysis methods and CNN architectures, revealing that SWT effectively represents TF textures of displacement signals. AlexNet, VggNet, and ResNet achieve identification accuracies of 83.55%, 92.76%, and 97.88%, respectively, with ResNet demonstrating the strongest identification capability. This method's primary contribution is validating the feasibility of utilizing displacement signals and CNN for slug, stratified, and elongated flow regime identification. Finally, SHAP analysis showed that minimum and maximum values extracted from SWT scalogram using ResNet architecture are the best flow-distinguishing feature using CNN.