High-accuracy recognition of gas–liquid two-phase flow patterns: A Flow–Hilbert–CNN hybrid model

Abstract

In multiphase flow pipelines, the timely and precise identification of flow pattern characteristics is crucial to guide industrial production, assess pipeline safety, and facilitate early risk warnings. Current research highlights that the combined application of flow imagery and convolutional neural networks (CNNs) can yield high accuracy flow pattern recognition. However, in the industrial sector, conveniently measurable signals such as pressure and flow rate present as one-dimensional signals rather than two-dimensional signals such as images and matrices, which CNNs are equipped to process. Therefore, the selection of a suitable data encapsulation type to convert one-dimensional signals into matrix or image form is vital to the successful application of CNNs in this field and to the precise identification of flow patterns in industrial production. In this research, we propose a Flow–Hilbert–CNN hybrid model, which integrates the one-dimensional liquid holdup signal, the Hilbert curve technique, and CNNs. The experimental data from multiphase flow in an undulating pipeline with low liquid holdup demonstrates that this hybrid model provides excellent classification accuracy, the model effectively identified four flow patterns—slug flow, pseudo-slug flow, wavy-stratified flow, and smooth-stratified flow—with recognition accuracies of 100.0%, 90.38%, 93.07%, and 100.0%, respectively, from a 16,384-length data source. Compared to models that employ standard data encapsulation methods (Space Folding (SF) and Fast Fourier Transform (FFT)) combination with CNNs, this hybrid model exhibits remarkable recognition accuracy and superior adaptability to data source scales. Notably, in the case of small-scale datasets, the accuracy improvement was 43.24% and 35.89% in comparison with SF- and FFT-based models, respectively. Moreover, the recognition accuracy of the hybrid model shows significant improvement when compared to results from traditional machine learning models and deep neural networks combined with the liquid holdup signal. The hybrid model, when used with different CNNs—MobileNetV2, Resnet18, and VGG16—yields recognition accuracies of 95.87%, 97.76%, and 96.64%, respectively, while the accuracies of machine learning models—deep neural network, k-nearest neighbors algorithm, and support vector machine—stand at 91.35%, 74.12%, and 90.57%, respectively. These results convincingly demonstrate the universality of combining the Hilbert curve with CNNs in the hybrid model, as well as the model's reliability and superiority in industrial application. This hybrid model can facilitate online, high-precision monitoring of flow patterns in the industrial sector and aid in preventing risk accidents within the production process.