A Framework for Flow Pattern Analysis and Identification Based on Dual-Domain Feature Extraction and Deep Learning

Abstract

A flow pattern analysis framework based on the combination of dual-domain feature extraction and deep learning is proposed. Measurement data of flow patterns are collected using a resistance sensor array. For flow pattern analysis, a novel approach combining Choi-Williams distribution (CWD) and limited penetrable visibility graph (LPVG) is proposed to construct complex networks. The average degree (AD) index and global efficiency (GE) index are calculated, and the topological structure of the complex network is analyzed to reveal the nonlinear dynamic behavior of flow patterns. Regarding flow pattern identification, a pseudo-image encoding (PIE) method is employed to encode the energy time series (Ets) from time-frequency analysis and the degree sequence from complex networks into two types of 2-D grayscale images to complete the feature extraction in the energy domain and the network domain. A novel deep learning classification model, the dual-input feature fusion network (DIFFN), is proposed to use two types of grayscale images as network inputs to complete flow pattern identification. The results indicate that our framework allows effectively characterizing the nonlinear dynamic behaviors during the evolution of different gas-liquid flow patterns. Meanwhile, the identification accuracy of the three flow patterns can reach 95%.