Structural health monitoring (SHM) techniques designed for single phase flow pipelines can be difficult to apply to multiphase flow pipelines. Data collected for pipeline SHM can be affected by intermittent changes in flow characteristics associated with flow patterns. Therefore, determining the flow pattern can help to better predict the unsteadiness of the flow parameters associated with pipeline integrity management. It is known that flow-induced vibration (FIV) in multiphase flow is highly correlated with flow pattern. In this paper, the characterization of FIV under various flow patterns is investigated experimentally, and a CNN-based flow pattern identification method is presented for horizontal gas-liquid pipe flow. Measurements were performed using two wall-mounted triaxial accelerometers with a high-speed camera, which simultaneously acquired both images and accelerometer signals. The difference of flow-induced vibration under various flow patterns can be explicitly shown by extracting morphological features using the Hilbert-Huang Transform (HHT) and Short Time Fourier Transform (STFT). In this paper, the HHT with 1st to 3rd IMFs is intended to emphasize the portion of FIV due to the unsteady fluctuations, while the STFT is intended to examine the FIV contributed to by both steady and unsteady fluctuations. To eliminate human interpretation errors, a convolutional neural network (CNN)-based machine learning model is built for the flow pattern identification task. To increase the size of the dataset used for flow pattern identification, the original time-series database is augmented by applying a series of data augmentation methods including sliding window, window stretch, denoising, and noise enhancement. These enable the flow pattern identification process to be more robust under various void fractions and flow geometries. The flow pattern identification results show that both HHT and STFT trained model have promising performance with overall accuracy above 97%. A comparison of identification results using HHT and STFT extracted morphological features from different accelerometer axes is performed. The results show that model trained by HHT images has a higher level of generalization.

This content is only available via PDF.
You do not currently have access to this content.